Podcast MCP
Provides audio processing, rendering, mixing, format conversion, and playback capabilities via FFmpeg.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Podcast MCP transcribe the host track and find long pauses"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Podcast MCP
FOSS podcast production toolkit: multitrack projects, per-track transcription, transcript-driven tightening, FFmpeg processing, ShareCut Studio (a full DAW-style editor in the browser and the desktop app), and agent skills under .agents/. The scope of the current beta is in docs/beta-plan.md.
Requirements: Python 3.11+, uv recommended (install.sh falls back to .venv + pip), and a matching FFmpeg/FFprobe 9.0.2 or later 9.x pair for source use. Desktop packaging selects a pinned prebuilt 9.0.2 pair. The release catalog pins the same platform archives for CI and desktop packaging.
Quick start
git clone <repo-url> && cd sharecut-studio
./install.sh
source .venv/bin/activate # or: uv run podcast …
podcast doctor
# Whisper defaults to large-v3-turbo (~1.6 GB, first transcribe). Smaller: ./install.sh --whisper-model small.en./install.sh installs a contributor set (dev + gui + relay). Optional downloads and extras: docs/setup.md.
Choose a downloaded Whisper model for one standalone transcription with podcast transcribe --project … --model small.en; use podcast setup --whisper-model small.en or PODCAST_WHISPER_MODEL to make that choice the machine default.
podcast transcribe reports the number of tracks processed by that run. When there are no dialogue tracks to process, it reports Transcribed 0 track(s). and warns on stderr.
No supported system pair? Use the prebuilt importer. Offline validation uses an explicit pinned archive through the same importer. Pair selection and the tested release policy, including process-local reuse of successful unchanged admission, are documented in docs/setup.md. Explicit health checks still run fresh bounded version probes.
Agent config is tool-agnostic under .agents/ (skills, rules, MCP template). Register MCP per docs/setup.md. With podcast gui running on loopback, Connect agent… (home or Menu) copies http://127.0.0.1:8765/mcp for URL-only clients. Use the podcast-setup skill when onboarding.
Codex UI edits have project-local Impeccable design hooks. Review and trust them in Codex Settings → Hooks after cloning or updating their definitions.
Related MCP server: AI Sound MCP Server
Maximal install (all extras)
Every pip extra, plus the models podcast bootstrap fetches: about 360 MB for the word aligner and about 80 MB for the speaker model. No extra pulls torch. Prefer the Quick start above for everyday CLI/MCP/GUI work.
git clone <repo-url> && cd sharecut-studio
uv sync --all-extras
source .venv/bin/activate # or: uv run …
podcast bootstrap
podcast doctor
# Optional ShareCut Studio UI (needs Node 24+):
cd gui/web && npm ci && npm run build && cd ../..Notes (verified in a disposable Linux Docker image):
uv sync --all-extrasinstalls every pip extra (dev,gui,relay,object-store,prosody).A bare
podcast bootstrapinstalls Whisper, RNNoise, the word aligner and the speaker model, and so does Studio's Get ready card.podcast bootstrap --component word-alignerinstalls just the aligner; once downloaded, precise word boundaries (transcribe.forced_alignment.enabled, English) turn on by default; until then they stay unavailable and the pipeline reports word times as Whisper-only (Studio: download it next to the Pipeline tab's Precise word boundaries field).Without
uv:python3 -m venv .venv && .venv/bin/pip install -e ".[all]"then the same bootstrap/doctor steps.
Detail: docs/setup.md § Maximal install. Operators can use the read-only public mirror check to verify bootstrap object availability without downloading payloads.
The frontend CI suite uses a locked exporter dependency group for its four generated-file checks. Browser suites keep the full GUI environment.
Repo layout
Path | Purpose |
| Python package — CLI, MCP, services, domain, GUI API, Extensions SPI |
| Optional private/provider extension, installed independently when needed; its source is not part of this repository |
| FOSS reverse-tunnel edge ( |
| ShareCut Studio viewer (React + Vite) |
| FOSS host-online share relay (Docker Compose examples) |
| Published caps / feature-id contract artifacts |
| Skills, rules, MCP template, pipeline defaults |
| Architecture and feature docs |
| Pytest suite and fixtures ( |
|
|
Episode workspace
podcast episode init --dir /path/to/my_episode
podcast track add --project /path/to/my_episode/episode.sharecut --id host --file raw/host.wav --role dialogue
podcast pipeline run --project /path/to/my_episode/episode.sharecutShareCut Studio (editor):
uv sync --extra gui
cd gui/web && npm install && npm run build && cd ../..
podcast gui --project /path/to/my_episode/episode.sharecutRelease wheels include the web build (the release-wheel workflow runs npm run build before uv build). A wheel built without it makes podcast gui warn at startup and / explain how to build it and set PODCAST_GUI_DIST (docs/setup.md). Hot reload: podcast gui --project … --dev --no-open plus cd gui/web && npm run dev. Agents: MCP open_gui_tool / skill podcast-open-gui (or podcast gui --background). Details: docs/gui-integration.md, docs/setup.md.
For the development-only UI component catalog, run cd gui/web && npm ci && npm run storybook; see docs/design-system.md for publishing setup.
A clean clone needs no relay, account, object store, CDN, or signing credentials for local editing. Before enabling collaboration or building a downstream desktop distribution, validate only that mode’s configuration:
podcast config check --mode local
podcast config check --mode self-hosted --relay-config config/relay.example.yaml
podcast config check --mode distributor --distribution-profile /path/to/distribution.jsonRuntime secrets belong in environment variables or
~/.config/podcast_mcp/relay.yaml; public desktop identity belongs in a validated
distribution profile. See docs/setup.md.
Edit transcription vocabulary with podcast transcript context set: pass
--project PATH --term NAME or remove entries with --remove-term NAME / --remove-guest-name NAME.
Host MCP agents can read and replace it with revision checks. See the
transcript workflow.
Documentation
ShareCut Studio Extensions — public plugin SPI (absent = no render); seams
Host-online relay — Docker share edge, tunnel, capability-scoped remote MCP (via the FOSS collaboration extension)
Share tokens and registry backups — create a new private backup with
podcast review backup-registry;--destrequires an unused name in an existing trusted private directory.Recording session (design) — record links, full-quality recording on each device, mix-minus monitor, consent
UX onboarding pack — shareable brief, screens, glossary, backlog (live site; See the UI)
UI philosophy — trust-first beta design principles for automation, recovery, and the free core workflow
Communication philosophy — voice, terminology, control placement, and the mobile ergonomics checklist for user-facing UI
Interaction system — one path per interaction (selection, targets, menus, sheets, feedback, keyboard and touch equivalents), the registries, and the checklist a new interaction passes
Setup — install, extras, optional downloads, MCP, bootstrap, play, GUI
Architecture — layers and timebase
Contributing — where new code goes; Git workflow (feature branch → PR → the base branch in
config/branches.json)Agent instructions — SOLID/DRY, tests-as-you-go, docs in sync
Episode format v2 — canonical
episode.sharecut(see also project-format.md)Multitrack ingest & alignment — raw files → one track per speaker
Transcript workflow — reconcile → precorrect → refine → audition
Session sync — agent ↔ DAW transport
GUI integration — viewer API, layers, live reload
Timeline comments — review feedback + action items (MCP/CLI/DAW)
CI workflow defines the frontend browser matrix and its job timeouts, including the Chromium and WebKit compatibility jobs.
Editor profiling:
npm --prefix gui/web run profile:remainingruns disposable workload profiles;profile:comparefreezes or applies local diagnostic budgets;make profile-seek-trim OUT=/tmp/seek-trimprofiles ruler seek and clip trim. See the profiling commands and limits.
Diagnostics and report-submission orchestration live in
services/support/; pipeline execution,
configuration, and component bootstrap live in
services/pipeline/. Contributor import
rules are in Contributing.
Support
Bugs: GitHub Issues. Attach a sanitized diagnostics zip from Home → ShareCut app menu → Help → Report a problem → Export or podcast doctor --bundle (writes ~/Downloads/sharecut-diagnostics-<UTC>-<nonce>.zip). Nothing is uploaded automatically — there is no telemetry. Do not attach episode audio or project JSON. Minted shares may transit the relay; GitHub issue attachments are public. Detail: docs/setup.md § Troubleshooting.
Privacy
Beta policy (local audio unless you mint a share; relay is a pass-through that does not store audio but share audio/media may transit while the share is active; no telemetry): sudo.science/privacy.html.
Play audio (listen while developing)
podcast play --project tests/fixtures/aligned_dialogue/episode.sharecut \
--source processed:reference --start 0 --end 15 # edits + FX (stem or segment render)
podcast play --project ... --source track:guest --start 0 --end 15 # raw alignment
podcast play --project ... --compare --start 18 --end 28 # alignment check
podcast play --project ... --query documented # after transcript is seeded
podcast play --project ... --dry-run # extract WAV only; see "tier" in JSONTesting
Tests run with a 95% minimum coverage gate (see docs/testing.md):
make test # full suite + coverage gate (parallel)
make test-fast # quick inner loop, no coverage
make e2e # aligned_dialogue fixture; see docs/e2e-fixture-manual.mdGUI frontend unit tests (session dedupe, etc.): cd gui/web && npm test.
GUI browser tests use npm --prefix gui/web run test:e2e; its wrapper owns private
registry and identity state. make ux-demo-screens captures host and guest views
from a disposable copy of the existing showcase fixture.
License
ShareCut Studio is licensed under the GNU Affero General Public License, version 3 or later (AGPL-3.0-or-later). See LICENSE.
Available Tools
218 toolsaccept_session_suggestion_toolA
Place a pending session on the timeline (one undo step, "Added Session N").
An insert opens the session's length at its insert point on every dialogue track and places its takes, room-tone beds and live comments there. A replace removes its range on every dialogue track and puts one take in its place (take_index, else the selected or last take) and discards the other takes' staged segments and the live comments held on them (the result's discarded_takes, comments_not_placed and discard_message say what; one Undo brings them back; the recordings stay in the project). Its range is found again from the audio it was staged on. Refuses a session that is not pending, a range or insert point that is gone or changed (stale_anchor), an unknown take (take_not_found), a take_index on an insert (take_not_applicable), a track the take did not record (track_not_recorded) and pending edits inside the range (pending_edits_in_range). A replace that also removes another track's speech stays unchanged and returns needs_confirmation until you repeat with override_audio_checks=true (recorded). The result lists non-dialogue tracks that did not move (unmoved_track_ids).
Inputs: suggestion_id (required), take_index, project_path, override_audio_checks. Returns: result
add_chapter_toolC
Add a chapter marker at a timeline time.
Inputs: title (required), at_time (required), project_path. Returns: result
add_comment_action_toolC
Append an action item TODO to an existing comment.
Inputs: text (required), comment_id (required), project_path. Returns: result
add_comment_reply_toolC
Append a flat reply under an existing timeline comment (threaded discussion).
Inputs: body (required), author (required), comment_id (required), project_path. Returns: result
add_comment_toolA
Add a timeline review comment (instant or span).
Times are timeline seconds (session/deliverable clock), like chapters - not source-media time. Use search_transcript_tool's timeline_start/end when anchoring to dialogue. track_ids / action_texts are optional string lists. Empty track_ids = session-wide. Undoable via history. Pass edit_decision_id to open the single Ask thread for a pending cut (unique per decision; later notes must use add_comment_reply_tool).
Inputs: body (required), author (required), timeline_start (required), track_ids, +4 more. Returns: result
add_effect_toolC
Add an effect (or preset) to a track or speaker's effect chain.
Inputs: effect, params, preset, speaker, +2 more. Returns: result
align_brief_toolB
Return the alignment acceptance brief: what the agent must check before marking done.
Inputs: project_path. Returns: result
align_done_toolC
Mark the conversation-clock alignment acceptance gate done, with optional notes.
Inputs: notes, project_path. Returns: result
align_retained_bleed_toolB
Align supported complete direct phrases locally while keeping uncertain audio intact.
Preview with apply=False. Saved manual/declined decisions take priority. Legacy placement overrides require an explicit timeline window.
Inputs: apply, end_sec, track_id, start_sec, +2 more. Returns: result
align_status_toolB
Report the conversation-clock alignment acceptance gate status (pending/done/waived).
Inputs: project_path. Returns: result
align_waive_toolC
Waive the alignment acceptance gate with a recorded reason.
Inputs: reason (required), project_path. Returns: result
analyze_cleanup_toolC
Analyze a track for cleanup opportunities (silence, clicks, hum).
Inputs: speaker, track_id, project_path. Returns: result
apply_bleed_suppression_toolC
Suppress mic-bleed words (dry_run previews without applying).
Inputs: apply, words, dry_run, end_sec, +5 more. Returns: result
apply_edit_plan_toolC
Stage cuts from an edit-plan list; applied entries undergo fresh cut assessment.
Inputs: edits (required), project_path, review_required, use_inaudible_opt. Returns: result
apply_editsC
Auto-apply non-review REMOVE/MUTE decisions whose reason starts filler: or pause:.
Inputs: project_path. Returns: result
apply_fade_recommendations_toolC
Apply a list of fade recommendations (from recommend_fades_tool) to the project.
Inputs: recommendations (required), project_path. Returns: result
apply_low_audibility_suppression_toolC
Suppress low-audibility transcript words, optionally only words; leaves audio unchanged.
Inputs: words, speaker, track_id, project_path. Returns: result
apply_suggested_actions_toolA
Apply every pending sound's own suggested action in one undo step.
Each tighten proposal carries a label and a suggested_action (propose_edits
shows them). This applies all of them, or only ids: cut and shorten cut
(a cut held by the automatic audio checks stays pending and is named in held),
lower turns a laugh down while another speaker talks over it (otherwise it is left),
leave dismisses the proposal and keeps the audio. overrides maps a decision id to
another action for that sound. Returns applied by action, cut_over_speaker,
held (ids still pending) and a summary sentence.
When the user says to apply the held ones anyway, repeat with
override_audio_checks=true (recorded in History and the edit log).
To change one sound after the batch (say a cut to leave) without undoing the rest, call
again with revise=true and overrides={id: action} and no ids: it steps back
to before the latest batch (which must still be the latest edit) and applies it again
with that change, as one undo step. Otherwise it refuses with
no_suggested_batch_to_revise.
Applying all pending sounds, or several ids, is a bulk apply: a cut or shorten
over a word the transcript is unsure about (below analysis.unsure_word_confidence, 0.7
as shipped) stays pending and is named in held and in held_unsure; when every cut is
held nothing is saved. Check those words (low_confidence_words_tool), then apply that
one cut by passing its single id in ids; override_audio_checks does not apply to it.
Inputs: ids, revise, overrides, project_path, +1 more. Returns: result
apply_transcript_cleanup_toolB
Batch word + phrase fixes in one undoable history step.
corrections: {"words": [{word_index, text}], "phrases": [{start_word_index, end_word_index, text}]}
Inputs: track_id (required), corrections (required), project_path. Returns: result
apply_transcript_gate_toolB
Apply an acoustic bleed plan to fresh dialogue stems and persist its source scope.
Retained speech and unresolved audio stay open. Later renders reuse the plan's policy and scope. Supported local retained-bleed alignment runs by default and respects saved timing choices. Alignment can trigger normal transcript reconciliation.
Inputs: apply, dry_run, end_sec, speaker, +5 more. Returns: result
approve_edits_toolA
Approve pending edits by id (or every pending proposal), applying them to the timeline.
To give every pending sound its own suggested action at once, use
apply_suggested_actions_tool.
all_pending=true approves every pending proposal; ids approves those. Either way
the edits that pass the automatic audio-quality and cut-position checks apply. ids lists
the edits this call approved, and held lists the ones stopped (each with its finding) and
left pending. unknown_ids lists requested ids that match no pending or approved edit (the rest still
apply); the call fails with approve_unknown_ids when none of the ids match. When the
user says to approve a held cut anyway, repeat with
override_audio_checks=true: it skips those checks for this call and records the
override (actor and surface) in History and the edit log.
More than one edit (all_pending or several ids) is a bulk approve: a cut that
covers a word the transcript is unsure about (below analysis.unsure_word_confidence,
0.7 as shipped) stays pending and comes back in held with reason unsure_word.
Check those words (low_confidence_words_tool), then approve that one cut by passing
its single id; override_audio_checks does not apply to it.
Inputs: ids, all_pending, project_path, override_audio_checks. Returns: result
approve_social_clips_toolC
Approve social clips by id.
Inputs: ids (required), project_path. Returns: result
audibility_map_toolC
Return per-word audibility scores for a track or speaker.
Inputs: speaker, track_id, project_path. Returns: result
audio_diagnostics_toolC
Run audio diagnostics (clipping, hum, level) on a time window.
Inputs: end_sec, speaker, track_id, start_sec, +1 more. Returns: result
audition_context_toolA
Caption each dialogue track in a timeline window + clip-skew / render freshness.
detail: summary (default) includes comments, active effects, edits, and
windowed hum/clip hypotheses; full expands edit/comment payloads;
visual adds waveform/spectrogram PNGs (slower). Returns
audition_context.v2 (typed hypotheses, suggested_listen, explicit clocks).
Each track also carries a prosody window (pitch, rate, energy, prominent
words, phrase boundaries) read from the pipeline's cached profile
(analyze_prosody step); status is missing/stale with a hint
when no fresh profile is cached yet, or unavailable (redacted error)
when the cache cannot be read. Top-level prosody_notes has up to
8 one-line summaries. Times are timeline (session) seconds. Does not play audio.
Inputs: end_sec (required), start_sec (required), detail, project_path, +1 more. Returns: result
bleed_words_toolC
List words contaminated by another speaker's microphone bleed.
Inputs: end_sec, speaker, track_id, start_sec, +1 more. Returns: result
bounce_audio_toolA
Bounce selected (or all) stems to export/bounces/ without mastering.
Times are timeline seconds. track_ids lists the stems, or omit for all
non-muted mixable tracks. formats lists extensions (e.g. ["wav", "mp3"]);
default wav only.
Inputs: end_s, formats, start_s, track_ids, +1 more. Returns: result
build_edit_contextB
Build the JSON edit-planning context: utterances with word indexes for cut_* tools.
Inputs: project_path, max_utterances. Returns: result
check_loudness_toolC
Measure LUFS and report artifact and dialogue-balance staleness without refreshing.
pass judges loudness; stale and stale_reason judge known project
artifact age. balance reports per-dialogue-track measurement freshness for
tracked project audio. An explicit unrelated file has stale=None.
Inputs: audio_path, project_path. Returns: result
clear_extracted_room_tone_bed_toolA
Take the extracted room-tone bed off track_id, in one undo step.
Only a bed cut from the track's own audio goes; a recorded or assigned bed is never
removed (cleared: false, reason: recorded_bed). Fills baked into mutes keep their
own copy of the bed; fills under timing holes follow the track's bed. Undo brings the bed
back. The clear lasts only until the next pipeline run with room_tone.extract_bed on,
which cuts a bed for a bedless track again.
Inputs: track_id (required), project_path. Returns: result
copy_segment_toolA
Copy [start, end) timeline seconds the way Studio Copy / Cut fill the clipboard.
Read-only. Returns {duration, extracts}: the clips of every track in range, or
only track_ids. Pass it to paste_segment_tool, which takes
the edit mode. To cut and paste, copy first, then cut (propose_range_cut_tool, or
delete_clips_tool with its mode).
Inputs: end (required), start (required), track_ids, project_path. Returns: result
correct_transcript_phrase_toolB
Replace a word range's text with corrected phrase text.
Pass expected_text (the space-joined words you read for this range) to refuse the fix
if those words changed meanwhile; omit to skip.
source_id names the recording whose words word_index counts; it is required on a
track recorded as several takes, and the refusal is word_source_required.
Inputs: new_text (required), track_id (required), end_word_index (required), start_word_index (required), +3 more. Returns: result
correct_transcript_toolA
Fix the text of one transcript word without changing its timing.
Pass expected_text (the word text you read) to refuse the fix if the word at
word_index changed meanwhile (for example a GUI edit shifted indices); omit to skip.
source_id names the recording whose words word_index counts; it is required on a
track recorded as several takes, and the refusal is word_source_required.
Inputs: new_text (required), track_id (required), word_index (required), source_id, +2 more. Returns: result
create_record_room_toolA
Mint a recording room: guest + producer /rec/ links; revokes earlier rooms' links (docs/recording-session.md).
If the current room has a stopped take not saved to the episode, the result says so (unlanded_take, unlanded_takes, unlanded_saving, unlanded_stalled, land_first_possible, warning) and the mint still happens. Pass land_first to save that take first (result adds landed_first; audio that never finished uploading cannot be saved; every mint that leaves audio the server detected as never finished reports left_behind (a count) and left_behind_message). Land-first mints nothing and keeps the old links when uploads are still saving (wait until 10 minutes pass with no upload and no rejoin from that person since Stop, or mint without it), when the room's staged session was already accepted or skipped and something is landable or saving, or when the land leaves a take unsaved.
target is where the session lands: end (the default), start (the front of the episode), or an insert/replace JSON object
that stages the session as a suggestion.
Inputs: target, expires_at, land_first, project_path, +1 more. Returns: result
crossfade_joins_toolC
Add crossfades at edit joins on a track, bounded by the clips only; dry_run previews.
Inputs: dry_run, fade_ms, speaker, track_id, +1 more. Returns: result
cut_text_match_toolB
Stage a pending cut for the first text match, or every match with match_all=True.
Inputs: query (required), speaker, track_id, match_all, +3 more. Returns: result
cut_time_range_toolB
Stage a pending cut of a time range on a track; applies after approve_edits.
Inputs: end (required), start (required), track_id (required), reason, +3 more. Returns: result
cut_utterance_toolC
Stage a pending cut of one utterance by its index.
Inputs: utterance_index (required), project_path, review_required, use_inaudible_opt. Returns: result
cut_words_toolC
Stage a pending cut of a word-index range on a track.
Inputs: track_id (required), end_word_index (required), start_word_index (required), source_id, +3 more. Returns: result
delete_clips_toolA
Delete whole clips by id, as the Studio clip Delete and Ripple delete do.
clip_ids lists the clips to delete (see list_clips_tool).
mode="gap" leaves silence where each clip was and moves nothing else;
mode="ripple" closes the clips' spans on every dialogue track, so speakers stay
in sync. A ripple that would cut speech outside the deleted clips (another track's,
or an unselected clip's on the same track) changes nothing and returns
needs_confirmation; repeat with override_audio_checks=true to apply it
anyway (recorded). Submits DeleteClip on the document plane; undoable.
For a time range across tracks use ripple_delete_tool.
Inputs: clip_ids (required), mode, project_path, override_audio_checks. Returns: result
delete_comment_toolC
Delete a timeline review comment (undoable via history).
Inputs: comment_id (required), project_path. Returns: result
document_command_schema_toolA
List the document command types, or return one type's payload JSON schema (read only).
With no command_type the result lists every type with submittable (false when
submit_document_command_tool refuses it, with the refusal naming the named tool to
use), the least guest_role and any host_only_reason. With a type it adds the
payload_schema to build the payload from. Time fields follow the command's own
payload: read each field's description before you send it. A baseline field
(SetEnvelope.expected_points, SetClipFade.expected, TrimClipEdge.expected_token)
says there how to read it.
Inputs: command_type. Returns: result
duplicate_segment_toolC
Duplicate a timeline segment at a new time.
Inputs: insert_at (required), source_end (required), source_start (required), project_path. Returns: result
edit_impact_report_toolC
Report pending-review and applied edits, with seconds cut in total and by track.
Inputs: markdown, project_path. Returns: result
episode_createC
Create a new episode project in a workspace directory and return its path.
Inputs: workspace_dir (required), name. Returns: result
export_audio_toolB
Encode mastered audio to export/ using pipeline.yaml formats, or formats objects.
All or nothing: files replace export/ only after every one is written. When the client
cancels the request (notifications/cancelled) the encode stops, the tool raises
CancelledProgress("Export cancelled") and an earlier export is left untouched. Do not
wait for a reply to the cancelled request (over JSON-RPC there is none); check export/.
Inputs: formats, project_path. Returns: result
export_social_clips_toolB
Export specified clips, or all approved and non-review-required drafts when IDs are omitted.
Inputs: ids, project_path. Returns: result
export_transcriptA
Write the edited episode's transcript next to the audio; return JSON.
formats: any of "txt", "md", "readable.txt", "srt", "vtt" (default ["md", "srt"]). Files go to the project's export folder under the episode name, all written together or none, and reflect the edited timeline (cut material omitted). dry_run writes nothing and returns the preview only; with_text adds the full reading text. Result keys: language, turns, duration_sec, preview, cue_sample, files (format, path, name, bytes, language, written, current), text. Errors: export_busy (another export is running), transcript_empty.
Inputs: dry_run, formats, with_text, project_path. Returns: result
extract_room_tone_bed_toolA
Cut a room-tone bed for track_id from its own quiet stretch, in one undo step.
The bed is what later fills (tighten pads, gap fills, bleed mutes) use in place of comfort
noise. A track with a recorded or assigned bed, or whose audio is unchanged since the last
extraction, is left as it is. With no clean stretch the reply says why
(extracted: false plus reason). Undo restores the earlier bed state. A later
clear_extracted_room_tone_bed_tool lasts only until the next pipeline run with
room_tone.extract_bed on, which cuts a bed for a bedless track again.
Inputs: track_id (required), project_path. Returns: result
fade_joins_toolA
Add fades at edit joins on a track; dry_run previews without applying.
Without fade_ms each join gets its own recommended fade, up to the episode's Longest
automatic join fade (set_join_fade_limit_tool), so joins can differ and come out
shorter than the limit. With fade_ms every join gets that length, still held to the
limit. The result reports fade_ms (the length applied; the longest, when joins differ),
requested_fade_ms, cap_ms (the limit in force) and kept_longer_joins (joins
that keep a longer fade already there). Fades you set yourself with set_clip_fade_tool
can be longer.
Inputs: dry_run, fade_ms, speaker, track_id, +1 more. Returns: result
feed_readA
Read the episode's podcast listing: every fact with where it came from, and what is missing.
facts (default) is JSON: each value says whether it is derived, inherited, measured,
minted, pinned (by a person, an agent or a guest, and whether the automatic value moved since) or
pending with the edit that would fix it, plus findings tagged blocker, warn or info.
notes_html is the show notes, item_xml the bare <item>, feed_xml the feed
file and show which show the episode belongs to. Use this before feed_set. Not
list_chapters_tool: chapters are edited there, never in the listing.
Inputs: view, project_path. Returns: result
feed_setA
Pin one podcast-listing value, as an agent. An episode value undoes with history.
item.* keys are pinned on this episode. channel.* keys and show (link this
episode to a show by name, creating it when new) change the show every episode shares,
are saved in the show's file and are not undone by history.
field is a key from feed_read, such as item.intro, item.title,
item.notes, item.enclosure.url, item.season or item.person[Name].href.
Write the episode intro with item.intro: one or two sentences, plain text. Setting
item.notes pins the whole notes text, so prefer item.intro. Blank text
unsets. value is text; a number or true/false is read as its text, so 2 and "2" both
set season 2. Chapters, clips, speakers and transcripts are edited where they live, not here.
Returns the updated facts. Not add_chapter_tool.
Inputs: field (required), value (required), project_path. Returns: result
feed_unsetB
Drop this episode's pin on a listing value so it goes back to automatic.
Also how to take the new automatic value after feed_read reports a pin whose
source changed. Unsetting a field that holds no pin changes nothing. An episode value undoes
with history; a show value (channel.*) is dropped from the show and does not.
Inputs: field (required), project_path. Returns: result
feed_writeA
Write the episode's feed file (<name>.feed.xml) next to its exported audio.
Refused while feed_read shows a blocker, such as no show, no hosting address or no
exported MP3. Idempotent: a second call with nothing changed writes nothing. The export
job also writes it. Returns the path, the files written and the file state.
Inputs: project_path. Returns: result
fill_with_room_tone_toolA
Fill a track's silences with its own room tone.
With track_id or speaker and no range the whole track's gaps are filled.
With start_sec and end_sec (timeline seconds, plus track_id) only the
clip-free time inside that range is filled; clips are never touched. A whole-track
fill on fewer than two clips replies unchanged.
Inputs: end_sec, speaker, track_id, start_sec, +1 more. Returns: result
find_replace_transcript_toolA
Replace a literal word or phrase everywhere in the transcript (Studio Find and replace).
Matches whole words across every track and recording; punctuation around a word is
kept. dry_run returns the preview (matches, count, skipped_words)
and changes nothing. Otherwise every previewed match is replaced through
ReplaceTranscriptMatches in one undo step, and replaced gives the count.
only is the 0-based position in the preview's matches of the one match to
replace (the Find bar's Replace); it cannot be combined with dry_run.
For one word use correct_transcript_tool.
Inputs: search (required), replacement (required), only, dry_run, +2 more. Returns: result
flagged_words_toolC
List words flagged for review (suppressed, low-audibility, or bleed).
Inputs: speaker, track_id, project_path. Returns: result
gate_overreach_toolC
Report where noise gating muted speech on a track.
Inputs: speaker, track_id, project_path. Returns: result
get_comment_toolB
Fetch one timeline comment by id (full action_items and resolve metadata).
Inputs: comment_id (required), project_path. Returns: result
get_session_presence_toolB
Who is in the ShareCut Studio session: display names, cursor, selection, viewport, transport, follow relationships.
viewport is the time window a follower should show. Near the start of a
phone's fixed-playhead timeline it begins at 0 and can be wider than what
is on that screen.
Inputs: project_path. Returns: result
get_session_state_toolB
Read shared DAW/agent session state (playhead, region, mode, selection).
Inputs: project_path. Returns: result
get_transcriptC
Fetch the transcript, combined or per-track, in text or JSON format.
Inputs: format, combined, project_path. Returns: result
get_transcript_vocabulary_toolC
Read project terms, guest names, revision, and re-transcription status.
Inputs: project_path. Returns: result
hear_session_suggestion_toolA
Hear a pending session before accepting it: the place as it is, after accepting, or both.
mode: current (the window as it is), suggested (the same window after accept_session_suggestion_tool, take_index chosen as there) or ab (current, a gap, then suggested; the default). pad_sec is the audio kept each side of the change (0 to 10 s, else pad_sec_invalid) and gap_sec the silence between A and B (0 to 5 s, else gap_sec_invalid). Nothing in the project, its history or the session changes, and a replace that would remove another track's speech still plays; the result's speech names the words. Refuses what accepting refuses, with the same codes: a session that is not pending, a range or insert point that changed (stale_anchor), an unknown take (take_not_found), a take_index on an insert (take_not_applicable), a track the take did not record (track_not_recorded) and pending edits inside the range (pending_edits_in_range). dry_run renders the WAVs without playing.
Inputs: suggestion_id (required), mode, dry_run, gap_sec, +4 more. Returns: result
history_diff_toolC
Diff two history indices, showing what changed between them.
Inputs: to_index, from_index, project_path. Returns: result
history_goto_toolC
Jump the project state to a history index.
Inputs: index (required), rerender, project_path. Returns: result
history_listC
List the project's history entries (index, label, mutation type).
Inputs: project_path. Returns: result
history_recordB
Record a named history checkpoint for the current project state.
Inputs: label, project_path. Returns: result
history_redoB
Redo the last undone project mutation.
expected_head_id guards it like history_undo (error_code history_stale).
Inputs: rerender, project_path, expected_head_id. Returns: result
history_status_toolB
Report history cursor, total, undo/redo availability, and current label.
Inputs: project_path. Returns: result
history_undoA
Undo the last project mutation (timeline, transcript, or settings).
Pass expected_head_id (history_status_tool's head_id) to undo only the change you looked at: if another edit landed since, the call refuses with error_code history_stale and nothing changes.
Inputs: rerender, project_path, expected_head_id. Returns: result
ingest_import_folder_toolC
Scan a recorder export folder and write ingest.yaml (audio-only, no copy).
Inputs: audio_dir (required), dry_run, out_manifest, speakers_override. Returns: result
ingest_suggest_alignment_toolC
Propose session_start offsets for multitrack ingest before consolidate.
Inputs: audio_dir (required), manifest_path (required), diag_dir, sweep_step, +5 more. Returns: result
ingest_verify_alignment_toolC
Audit consolidated dialogue tracks with VAD overlap on a timeline window.
Inputs: diag_dir, project_path, window_end_sec, write_waveforms, +1 more. Returns: result
insert_gap_toolB
Open pure silence on the timeline (split + shift).
Do not use to fake a beat after a tight handoff unless the user asks —
prefer keeping existing room tone via suggest_handoff_cut_tool bounds.
Inputs: at_time (required), duration_sec, project_path. Returns: result
inspect_clip_bleed_toolC
Inspect automatic attenuation on one occurrence and return guarded Restore targets.
Inputs: clip_id (required), track_id (required), project_path. Returns: result
join_label_toolC
Record a pass/fail label for a join to train the join-quality ranker.
Inputs: verdict (required), join_sec (required), note, play, +6 more. Returns: result
join_qa_sweep_toolC
Score all joins on a track (or the whole project) for QA review.
Inputs: track_id, project_path. Returns: result
join_quality_toolC
Score an existing or proposed cut join for audibility artifacts (clicks, glitches).
Inputs: cut_end, speaker, join_sec, timebase, +3 more. Returns: result
kit_apply_toolA
Put every segment of a kit on an episode, in order (podcast kit apply).
Each entry is an InsertSegment on the document plane (journaled, shown in Studio, later
content ripples), so end means the end after the earlier entries. There is no compound
command: every entry is its own undo step, undo_steps in the result says how many.
Use dry_run=true first: it runs the whole kit on a copy, reports each entry's effect
and the new duration, and changes nothing (not even a legacy project rename). An unknown
kit, a segment or pinned version missing from the library, changed segment audio and any
refusal the copy run finds are reported before anything is written. A failure of the
write itself (disk, lock, conflict) can still stop the apply after earlier entries went
in; the error then names the entry and how many were applied, and each applied entry is
its own undo step. Errors: kit_unknown, kit_segment_missing,
segment_hash_mismatch, segment_position_out_of_range, segment_media_invalid,
segment_write_failed.
Inputs: kit_id (required), dry_run, project_path. Returns: result
kit_list_toolA
List the show kits: the newest version of each, with how many of its segments are missing.
A kit whose newest version is damaged is listed as {"id", "version", "damaged": true, "problem", "code"}. Read-only; kits are per user, like the segment library.
Inputs: no named parameters. Returns: result
kit_save_toolA
Save a show kit (podcast kit save): library segments in order, each with where it goes.
Each entry is {"segment_id", "at", "version"?, "note"?}; at is start, end
or a number of seconds (read when the kit is applied), version pins a segment version
(omitted = the newest when the kit is applied; not a whole number from 1 is
segment_version_invalid, as on the CLI). The same name with other entries becomes
the next version; the same entries are reported unchanged. Returns
{status, id, version}. Errors: kit_name_invalid, kit_invalid,
kit_segment_missing (names every entry whose segment is not in the library).
Inputs: name (required), entries (required). Returns: result
kit_show_toolB
Show one kit: its segments in order, where each goes, and whether each is in the library.
Errors: kit_unknown, kit_unreadable.
Inputs: kit_id (required). Returns: result
list_applied_edits_toolC
List applied edits in a timeline window.
Inputs: track_id, project_path, timeline_end, timeline_start. Returns: result
list_chapters_toolC
List chapter markers in timeline order.
Inputs: project_path. Returns: result
list_clips_toolA
List clips in timeline order for one track or all tracks when track_id is omitted.
A clip with timing_steps has timing corrections, so its timeline length is not
source_end - source_start; include_timing_steps adds the step list.
Inputs: track_id, project_path, include_timing_steps. Returns: result
list_comments_toolA
List timeline review comments sorted by timeline_start.
Pass include_resolved=false for an agent work queue, or open_actions_only=true for incomplete action items. Each row includes id, body, author, track_ids, action_items, resolved_*.
Inputs: project_path, include_resolved, open_actions_only. Returns: result
list_edit_decisions_toolB
List pending decisions with Review fields; archived edits are in list_applied_edits_tool.
Inputs: applied, project_path, reason_prefix, review_required. Returns: result
list_effects_toolC
List the effect chain for a track or speaker.
Inputs: speaker, track_id, project_path. Returns: result
list_exported_transcriptsB
List transcript files already exported next to the audio as JSON.
Each entry: format, path, name, mime, language, bytes, modified (UTC), current (false when the transcript was edited since the file was written). Read-only.
Inputs: project_path. Returns: result
list_review_versions_toolC
List frozen review mix versions (includes active flag).
Inputs: project_path. Returns: result
list_session_suggestions_toolA
List record sessions staged for an insert or replace target.
status: pending, skipped or accepted; empty lists all. A staged session has its recordings registered as sources but nothing on the timeline until accept_session_suggestion_tool places it.
Inputs: status, project_path. Returns: result
list_session_takes_toolA
List the takes of a staged session for a take switcher (reads only).
Per take: take_index, duration_s (and per track), word_count and speech_text (a snippet; null for a take nobody has transcribed), chosen (the one select_session_take_tool picked), default (the one accept_session_suggestion_tool places when no take_index is given) and refusal ({code, message}) when accepting that take would be refused. selectable is false for an insert (its takes play in order), a session that is not pending and a replace with one take.
Inputs: suggestion_id (required), project_path. Returns: result
list_social_clips_toolC
List proposed social clips, optionally filtered by approved status.
Inputs: approved, project_path. Returns: result
low_audibility_words_toolC
List words too quiet to hear relative to the mix.
Inputs: speaker, track_id, project_path. Returns: result
low_confidence_words_toolB
List words below a confidence threshold for human review.
threshold defaults to the project's unsure-word line, the config key
analysis.unsure_word_confidence (0.7 as shipped), the same line the mapped unsure flag,
the refine brief and bulk approve use. Suppressed words are not listed.
Inputs: threshold, project_path. Returns: result
model_downloadA
Download one model, only after the user agrees: ask first, never call this on your own.
component is the model.id of a model_needed refusal (whisper, rnnoise,
word-aligner or speaker-ecapa); whisper_model picks the Whisper size and
upgrade re-fetches a model that no longer matches its pin. Inside the ShareCut Studio
process (this server reached through Studio's own MCP endpoint) it starts a download job
and returns its job_id: poll model_download_status. In any other server, including a
stdio server next to a running Studio (this tool is not handed to Studio, as it has no
project_path), it downloads in this call, which can take minutes for Whisper, reports
progress while it runs, and starts no job, so model_download_status answers
{"running": false} and Studio does not see it.
Inputs: component (required), upgrade, whisper_model. Returns: result
model_download_statusA
Report a model download started by model_download: status, current file and any failure.
Omit job_id for the latest download. Only the ShareCut Studio process has download jobs;
anywhere else (a stdio server, even with Studio running) the answer is {"running": false}.
Inputs: job_id. Returns: result
move_by_text_toolC
Move a transcript-matched segment before or after another transcript match.
Inputs: source_query (required), destination_query (required), position, project_path. Returns: result
move_clips_toolA
Move clips in session time and/or onto another track (no neighbor ripple).
clips is a list of {clip_id, timeline_start, track_id}.
Inter-track moves keep the originating media via source_id. Unlike
move_segment_tool, this does not cut a range on every dialogue lane.
Inputs: clips (required), project_path. Returns: result
move_segment_toolC
Cut a timeline segment and reinsert it at a new time (ripple move).
Inputs: insert_at (required), source_end (required), source_start (required), project_path. Returns: result
mute_off_speaker_toolA
Mute a track wherever another speaker is talking and its own speaker is not.
Use after transcript reconcile on a call recording whose track still carries another
person's voice. A stretch is muted when another dialogue track has words there, this
track has none, and this track holds nothing louder than a copy of that speaker.
Each stretch is a visible muted region in Studio (Restore brings it back); History
undoes the whole command. Preview with dry_run=true.
Inputs: apply, dry_run, speaker, track_id, +1 more. Returns: result
open_gui_toolA
Start the read-only DAW web viewer (background) and return its URL.
Idempotent: if a healthy viewer is already listening on host:port, reuses it
and optionally opens the browser to this project. Requires the gui extra
(uv sync --extra gui) and a built gui/web dist for the HTML UI.
Inputs: host, port, open_browser, project_path. Returns: result
overlap_duplicates_toolC
Find duplicate words across tracks in an overlap window.
Inputs: end_sec, start_sec, project_path. Returns: result
paste_segment_toolA
Paste a copy_segment_tool clipboard at insert_at timeline seconds (Studio Paste).
mode="ripple" opens duration of time on every dialogue track at
insert_at, then places each extract on its own track. mode="gap" pastes
over: it replaces [insert_at, insert_at + duration) on the pasted tracks only,
and nothing moves. The same clipboard can be pasted again. A clipboard naming an
unknown track or source, or an impossible range, is rejected with a paste_*
code. Submits PasteSegment on the document plane; undoable.
Inputs: clipboard (required), insert_at (required), mode, project_path. Returns: result
pipeline_analyze_toolC
Heuristic Analyze: propose pipeline param patches from audio diagnostics.
Inputs: apply, project_path. Returns: result
pipeline_get_config_toolB
Return effective pipeline config, step metadata, and param schema (same as GUI).
forced_alignment is the resolved Precise word boundaries state (enabled,
requested, installed, blocked, reason): on by default when the word
aligner is downloaded, unavailable until then (#780).
Inputs: project_path. Returns: result
pipeline_runB
Run the production pipeline. Optional skip_steps (step names) / config (object) override yaml.
When use_working_set is true (default), also reads/writes the GUI session
working set so agents and ShareCut Studio share the same visible config.
Omit unattended to leave the working-set Batch mode unchanged.
Existing transcripts are reused; force_transcribe=true re-runs ASR for this run
only (not persisted). retime_words=true re-times stored transcripts with the forced
aligner from their ASR cache for this run only (no Whisper; hand-edited transcripts are
skipped); check transcript_timing.json → forced_alignment.retime for skipped and
failed tracks; not with force_transcribe. It fails before changing any transcript
when the word aligner is not downloaded (the error names
podcast bootstrap --component word-aligner). Forced alignment is otherwise on by
default whenever that model is installed (transcribe.forced_alignment.enabled unset);
forced_alignment.reason in transcript_timing.json says what a run resolved to.
Replacing edited transcripts is refused when
unattended; run attended, or the user confirms with Studio Re-transcribe.
Returns Completed through <step>; when this run exported, the next lines are
the export QC verdict (same as CLI pipeline run).
Inputs: config, from_step, only_step, skip_steps, +5 more. Returns: result
pipeline_set_config_toolC
Update the shared GUI/agent pipeline working set (visible params).
Inputs: reset, config, unattended, project_path, +1 more. Returns: result
play_ab_toolA
Extract the same range at two history indices, then play A→gap→B once.
Prefer this over history_goto → play → goto → play for before/after probes.
Leaves the history cursor on after_index. Default gap is 0.4s.
Inputs: after_index (required), before_index (required), source, dry_run, +5 more. Returns: result
play_ab_wavs_toolB
Play two existing WAVs back-to-back with a short silence gap (one concat).
Inputs: wav_a (required), wav_b (required), dry_run, gap_sec, +1 more. Returns: result
play_audio_toolC
Extract and optionally play audio (processed:, track:, premix, export).
Inputs: raw, query, source, compare, +8 more. Returns: result
play_compare_toolC
Play each dialogue track then premix for the same timeline range.
Inputs: dry_run, end_sec, rerender, start_sec, +1 more. Returns: result
play_compose_toolB
Mix selected tracks for a timeline window and optionally play (no project mutation).
tier is processed (edits+FX stems) or raw (source media). Times are
timeline (session) seconds. Does not change mute/solo/FX on the project.
Inputs: end_sec (required), start_sec (required), track_ids (required), tier, +3 more. Returns: result
play_pending_preview_toolB
Hear Current vs Suggested vs A/B full mix for a pending remove or mute.
Suggested renders the window after approving the edit on a snapshot: the same
ripple, paced pad, fades and mute that approving ships. Splits have no Suggested
side. pad_sec is the audio either side of the edit, 0 to 10 seconds (default 0.5);
outside that range it is refused with pending_preview_pad_out_of_range.
Does not mutate the project.
Inputs: edit_id (required), mode, dry_run, gap_sec, +3 more. Returns: result
play_transcript_query_toolC
Search transcript for query and play that span (NL: play where they talk about X).
Inputs: query (required), dry_run, rerender, processed, +3 more. Returns: result
precorrect_transcript_toolB
Apply automatic transcript precorrections; dry_run previews without changing words.
Inputs: dry_run, project_path. Returns: result
preview_inaudible_cut_toolA
Dry-run local inaudible-cut snap (~±80ms) for a proposed range.
Not a multi-second narrative handoff planner. For leave-a-beat / clean-up-the-
transition / punchline-to-pivot bounds use suggest_handoff_cut_tool.
Inputs: end (required), start (required), speaker, timeline, +2 more. Returns: result
propose_editsA
Propose filler/pause tighten edits for review (does not apply them).
Returns a JSON object {operation, edits, skip_counts, summary} (not a
bare array — breaking vs older list-of-decisions clients). operation is
propose_edits. Each edit also carries label (plain words, e.g. "Laugh"),
kind (filler / pause / repeat / false-start / sound / other),
sound_kind (the raw kind, e.g. laugh; null for other)
and suggested_action (cut / shorten / lower / leave) from
one table (edits/cut_policy.py) and audio_check (clean, or held with the reason
Approve all would give, found on a scratch copy; absent when the audio could not be read).
skip_counts maps discourse:{token} to kept uses
and backchannel:{phrase} to kept acknowledgments (uh huh).
word_probe:unavailable is a missing Whisper model; word_probe:failed
is a span whose decode raised. The summary names both.
edit_mode is ripple (default, from tighten.edit_mode) or mute.
Mute proposes EditDecisionType.MUTE filler hits and skips pause
candidates (muting a pause is a no-op). A pause: trim is shrunk to the air inside
the pause, labelled :air_edges when it differs from the span pacing proposed, and
proposed only when it removes at least a tenth of its pause (imperceptible skips
the rest). It is reviewed or applied by the checks a filler gets. intensity is light /
medium / aggressive (default from tighten.intensity, else medium):
a deterministic preset over existing tighten keys (light = clustered um/uh
only and >=0.5 s pause air; medium also proposes a lone um/uh; aggressive =
isolated and borderline discourse markers, 0.3 s solo pauses). Unknown
values raise. Suggest a tier from the episode (interview -> light, solo
monologue -> aggressive); the pipeline itself never chooses. Pipeline
auto-tighten stays off; listen-first review before apply_edits /
approve_edits. Not NL cut-by-text (cut_* tools) or narrative focus (focus
tools).
Inputs: edit_mode, intensity, project_path. Returns: result
propose_range_cut_toolA
Propose an exact range cut for host review (Studio range Cut), replay-safe.
start / end are timeline seconds; track_ids lists the lanes
(default every dialogue lane). The cut leaves a hole: later clips keep their places.
This never applies audio: it creates a pending exact cut that the host auditions and
approves (podcast edit approve also can). Reuse command_id only when retrying the
same range. For a ripple cut across lanes use ripple_delete_tool.
Inputs: end (required), start (required), command_id (required), track_ids, +1 more. Returns: result
propose_range_mute_toolA
Propose an exact range mute for host review, with replay-safe command identity.
Pass a serialized ExactRangeTarget from bleed preview or the range editor. This never applies audio or approves the proposal. Reuse command_id only when retrying the same target. The host auditions and approves or rejects it.
Inputs: target (required), command_id (required), project_path. Returns: result
propose_social_clips_toolB
Propose short-form social clips for a platform, limited to max_clips.
A rerun keeps approved and manually added clips and replaces the rest.
Inputs: platform, max_clips, project_path. Returns: result
publish_review_version_toolA
Freeze the current premix/mastered mix as a review version under artifacts/review/.
Comments created while a version is active stamp review_version_id. prefer: "premix" (default) or "mastered". Refuses a project with no mix yet (code no_mix: Refresh first), a premix that's stale vs the project (Refresh first) or a master not mastered from the current premix (export first).
Inputs: label (required), prefer, set_active, project_path. Returns: result
recommend_fades_toolC
Recommend fades at joins to hide edit artifacts.
Inputs: speaker, track_id, project_path. Returns: result
reconcile_transcript_toolC
Re-run transcript/timeline reconciliation; dry_run previews without applying.
Inputs: dry_run, end_sec, speaker, track_id, +2 more. Returns: result
reconciliation_status_toolC
Report transcript/timeline reconciliation status (stale vs current).
Inputs: project_path. Returns: result
record_assign_track_toolA
Correct a person's track: put their recordings on an existing track id, or pass track="new" for a new one. Allowed until the session is accepted; one undo step when takes are already staged; a replace follows the person and returns to their old track once no staged take or bed of theirs is left, unless a staged take or bed sits on either track or another person holds it by choice or by an acknowledged, non-tombstoned take (the moved person's own pin or take does not block). Refusal codes: unknown_participant, unknown_track, track_taken, replace_target_track, person_removed, session_already_accepted, session_skipped.
Inputs: track (required), participant_id (required), project_path. Returns: result
record_discard_take_toolB
Delete a terminal take (refused while its full-quality recordings are still saving to the project).
Inputs: take_index (required), project_path. Returns: result
record_land_toolC
Land saved full-quality recordings and live comments on the timeline.
Inputs: project_path. Returns: result
record_marker_toolB
Host Marker: stamp the open take now; record_land_tool lands it as a comment.
Inputs: body, project_path. Returns: result
record_pause_toolC
Host Pause for the current take.
Inputs: project_path. Returns: result
record_resume_toolC
Host Resume for a paused take.
Inputs: project_path. Returns: result
record_room_tone_toolB
Host Room tone: tell every recorded device to capture one minute of room tone now (lobby or stopped only).
Inputs: project_path. Returns: result
record_start_toolB
Host Start for the current record room (a solo take works; blocked while a joined guest has not consented).
Inputs: project_path. Returns: result
record_state_toolB
Read the live recording-room snapshot (roster, take clock, consent, and the mint preflight: unlanded_take / unlanded_takes = stopped takes not yet saved to the episode, unlanded_saving = how many of those are still uploading or not started, unlanded_stalled = recordings that never finished uploading, land_first_possible = a prediction: true when a land could save a finished recording, nothing is saving and the room's session was not already accepted or skipped; false in such a decided room).
Inputs: project_path. Returns: result
record_stop_toolD
Host Stop for the current take.
Inputs: project_path. Returns: result
record_target_toolB
Host Target: set where this room's session lands (end, start, or an insert/replace JSON object that stages the session as a suggestion; refused once a take exists).
Inputs: target, project_path. Returns: result
record_tracks_toolC
Show which track each recorded person lands on (matched by name to an existing track, new, or ambiguous, which also covers a track already used by someone else: reason track_in_use), with a plain-words reason. Host only.
Inputs: project_path. Returns: result
reject_edits_toolB
Reject pending edits by id, discarding them without applying.
Inputs: ids (required), project_path. Returns: result
reject_social_clips_toolC
Reject social clips by id.
Inputs: ids (required), project_path. Returns: result
remove_chapter_toolC
Remove a chapter marker by title.
Inputs: title (required), project_path. Returns: result
remove_effect_toolC
Remove an effect from a track or speaker's effect chain.
Inputs: effect, speaker, track_id, project_path. Returns: result
render_finalC
Render the final mastered export and return its path.
Inputs: project_path. Returns: result
render_previewC
Render a preview mix and return its status and path.
Inputs: rerender, project_path. Returns: result
render_status_toolC
Report stem freshness and whether dialogue balance measurements are stale.
Inputs: project_path. Returns: result
resolve_comment_toolA
Mark a comment resolved (or reopen with resolved=false).
Always pass by= identity (e.g. "agent" or the human name). Records resolved_by / resolved_at. Prefer resolving after action items are done.
Inputs: by (required), comment_id (required), resolved, project_path. Returns: result
restore_clip_bleed_toolA
Save one inspected Restore choice. Reuse command_id when retrying the same submission.
Does not render audio or restore other occurrences. Changed targets must be inspected again.
Inputs: target (required), command_id, project_path. Returns: result
revert_applied_edit_toolB
Restore applied edits as one History step; dry_run plans without writing.
Returns a RestoreResult. Records with restore.state history_only need history_undo.
Inputs: record_ids (required), dry_run, project_path. Returns: result
revoke_record_room_toolC
End a recording room: revoke guest and producer tokens.
Inputs: session_id (required), project_path. Returns: result
ripple_delete_text_toolB
Ripple-delete the first timeline span matching a transcript text query.
The matched speaker's words are the cut; other tracks' speech in the span needs
confirmation.
The automatic audio checks hold a harsh or other-speaker cut and name the finding;
override_audio_checks=true applies it anyway and records the override.
Inputs: query (required), project_path, use_inaudible_opt, override_audio_checks. Returns: result
ripple_delete_toolA
Ripple-delete a timeline range across every dialogue track.
Default inaudible opt absorbs trailing quiet air to ~0.4s before the next
word. For punchline-to-pivot / leave-a-beat handoffs, call
suggest_handoff_cut_tool first and pass use_inaudible_opt=false so
local snap does not pull mid-silence bounds onto speech. track_ids
names the tracks whose speech you mean to cut; speech on any
other track in the range needs confirmation. Omitted, the cut names no track,
so speech on every track needs confirmation (a whole-session time cut).
The automatic audio-quality and cut-position checks run first: a cut that lands
mid-word, makes an audible join or cuts another track's speech stays unchanged and
comes back with the finding. When the user wants that exact cut anyway (or names the
position themselves), repeat with override_audio_checks=true; it keeps the bounds as
given and records who overrode in History and the edit log.
Inputs: end (required), start (required), track_ids, project_path, +2 more. Returns: result
roll_clip_join_toolA
Roll the join between two neighbouring clips by delta_sec source seconds.
Positive moves the join later: the left clip gains material and the right clip
starts later in its source, so the pair keeps its length and later clips stay put.
Same operation as the Studio roll seam (RollClipJoin); undoable. The clips must
abut (touch within the join tolerance, see clips_abut); across a gap the call is
refused with error_code roll_needs_abutting_clips and nothing changes. Use
trim_clip_edge_tool to move one edge and ripple instead.
Inputs: delta_sec (required), left_clip_id (required), right_clip_id (required), project_path. Returns: result
search_transcript_toolC
Search the transcript for text, optionally scoped to a track or speaker.
Inputs: query (required), speaker, track_id, project_path. Returns: result
seek_session_toolC
Seek the open DAW playhead (no OS audio). Optional selection highlights a modifier.
Inputs: playhead_sec (required), selection, project_path. Returns: result
segment_insert_toolA
Insert a library segment into an episode at start, end or a timeline second.
at is the text start or end (any case) or a number of seconds; end is read
when the edit runs, under the document lock.
Later clips, chapters, review comments and pending edits move by the segment's length
(podcast segment insert). The edit goes through the document plane as InsertSegment:
it is journaled, shows in Studio at once and is one undo step (history_undo). It is
applied, not held as a pending proposal, because a pending proposal only exists for
remove, mute and split edits. Use dry_run=true first: it reports the new duration and
what would move, and changes nothing (it also probes the segment's audio, so it refuses
what the insert would). version picks an older version; the one used is journaled.
The result carries dry_run (true or false). Errors: segment_version_invalid, segment_unknown, segment_hash_mismatch, segment_position_out_of_range,
segment_media_invalid, segment_write_failed.
Inputs: at (required), segment_id (required), dry_run, version, +1 more. Returns: result
segment_list_toolA
List the segment library: the newest version of each saved intro, outro, sponsor read or sting.
A segment whose newest version is damaged is listed as {"id", "version", "damaged": true, "problem", "code"} rather than hidden. Read-only; the library is per user, not per episode.
Inputs: no named parameters. Returns: result
segment_save_toolA
Save clips of an episode into the library as a copy (podcast segment save).
Pass clip_ids or time_range (such as 0:00-0:22), not both. kind is
intro, outro, sponsor, sting or other; it is guessed from the position when omitted. The
episode is not changed (a project still under its retired entry-file name is not renamed). The same name with new audio becomes the next version of that
segment; identical audio is reported unchanged. Returns {status, id, version}.
Inputs: name (required), kind, role, clip_ids, +2 more. Returns: result
segment_show_toolB
Show one library segment: layout, level, chapter title, clips and its version numbers.
Errors: segment_unknown (no such id), segment_hash_mismatch (its audio changed).
Inputs: segment_id (required). Returns: result
select_session_take_toolA
Pick the take a pending replace session will place (one undo step).
Accept and hear_session_suggestion_tool use it unless given another take_index. Refuses a session that is not pending (session_suggestion_not_pending), an unknown take (take_not_found), an insert (take_not_applicable), a track the take did not record (track_not_recorded), a range that changed (stale_anchor) and anything else accepting that take refuses. Choosing the take already chosen changes nothing.
Inputs: take_index (required), suggestion_id (required), project_path. Returns: result
set_active_review_version_toolC
Set or clear review.active_version_id (pass null/omit to clear).
Inputs: version_id, project_path. Returns: result
set_clip_fade_toolA
Set the fade-in and fade-out duration of a single clip, and optionally their shapes.
A fade you set is bounded only by the clip: each fade fits in the clip, and fade-out is
limited to what fade-in leaves. render.join_fade_max_ms and the episode's Longest automatic
join fade (set_join_fade_limit_tool) limit only the fades ShareCut picks by itself.
The result reports the applied fade_in_ms / fade_out_ms and clamped. A curve is linear (the default),
equal_power or s; left out, the clip keeps its current curve. A curve applies to a
fade that is not at a cut or blended join. list_clips_tool names a curve only while it
is not linear.
Inputs: clip_id (required), fade_in_ms (required), fade_out_ms (required), project_path, +2 more. Returns: result
set_clip_gain_toolA
Set the level of a single clip in dB (0 = unity, range -60 to +12).
Clip gain applies before the track chain, staging gain, and fader. A value outside the range is refused. Changing it makes the track's stem stale.
Inputs: clip_id (required), gain_db (required), project_path. Returns: result
set_clip_join_toolA
Set the join between two neighbouring clips: mode plus fades (TOOL_TIMEBASE: na).
mode is fade | crossfade | cut. Crossfade sets left fade-out and right fade-in to
length_ms (default tighten.crossfade_ms); cut zeroes both fades; fade sets both
(or keeps them). Prefer this over set_join_mode_tool, which changes the mode only
and leaves a crossfade with no fades to blend.
Inputs: mode (required), left_clip_id (required), right_clip_id (required), length_ms, +1 more. Returns: result
set_clips_fade_toolA
Add a delta (ms, negative shortens) to the fade-in and fade-out of several clips at once.
One undo step covers every clip. Each clip is bounded by its own length like
set_clip_fade_tool, not by the Longest automatic join fade. A curve (linear,
equal_power or s) is set on every named clip; left out, each clip keeps its own. A clip changed since it was read returns clip_fade_changed and no clip changes.
Inputs: clip_ids (required), project_path, fade_in_curve, fade_out_curve, +2 more. Returns: result
set_clips_gain_toolA
Set several clips to the same level in dB (absolute, 0 = unity, range -60 to +12).
One undo step covers every clip. A value outside the range is refused and no clip changes.
Inputs: gain_db (required), clip_ids (required), project_path. Returns: result
set_comment_action_done_toolB
Check or uncheck a comment action item; records completed_by / completed_at.
Agents should pass by="agent" (or a stable agent label) when finishing work.
Inputs: by (required), action_id (required), comment_id (required), done, +1 more. Returns: result
set_effect_bypass_toolC
Toggle bypass on one chain effect by index (TOOL_TIMEBASE: na).
Inputs: bypass (required), effect_index (required), speaker, track_id, +1 more. Returns: result
set_envelopeA
Set the volume automation envelope for a track from [{time, value}] points.
Submits a SetEnvelope document command (same lock, conflict check, undo
history, and GUI fanout as a DAW edit). expected_points is the
[{id, time, value}] list you last read; if the envelope changed since,
the call fails with a conflict instead of overwriting the edit. When
omitted, the envelope as read at call time is the baseline.
Inputs: points (required), track_id (required), project_path, expected_points. Returns: result
set_join_fade_limit_toolA
Set the episode's Longest automatic join fade in ms (default 40, 0 to 200); null returns to the app default.
This is a limit, not an applied length: ShareCut picks each join's fade itself
(fade_joins_tool, Smooth all joins via apply_fade_recommendations_tool, approve-time
fades) and never goes above it, so joins can come out shorter (a limit of 120 may give 77 ms).
To give joins an exact length use fade_joins_tool with fade_ms (also held to the
limit, so raise it first for longer joins). Fades set by hand with set_clip_fade_tool,
set_clips_fade_tool or set_clip_join_tool can be longer, and no existing fade changes. Saved in the episode file and undoable; it makes no stem
stale. list_clips_tool reports the value in force as join_fade_limit_ms.
Inputs: ms (required), project_path. Returns: result
set_join_mode_toolA
Set join_in_mode on one clip: fade | crossfade | cut (TOOL_TIMEBASE: na).
Mode only: fades are untouched, so a crossfade with no fades renders as a plain
join; the result's join_crossfade_blocked says why. Prefer set_clip_join_tool.
Inputs: clip_id (required), join_in_mode (required), project_path. Returns: result
set_retained_bleed_alignment_mode_toolB
Save a timing choice for an applied decision or a proposal from the same window.
Manual and declined choices prevent later automatic correction; auto resets that choice.
Inputs: mode (required), decision_id (required), end_sec, track_id, +3 more. Returns: result
set_session_mode_toolC
Set DAW audition mode: mix | fx | raw.
Inputs: mode (required), project_path. Returns: result
set_session_playing_toolB
Start or pause DAW browser transport (no OS audio).
Inputs: playing (required), project_path. Returns: result
set_session_region_toolC
Highlight a timeline region in the DAW; optionally start browser playback.
Inputs: end_sec (required), start_sec (required), query, playing, +2 more. Returns: result
set_session_selection_toolC
Highlight a DAW modifier: selection {kind, id?, track_id?, time?}; null or omit to clear.
Inputs: selection, project_path. Returns: result
set_transcript_vocabulary_toolA
Replace vocabulary using the revision returned by get_transcript_vocabulary_tool.
Pass null only when that read returned a null revision. A stale revision rejects the write; the existing show title and other transcript context stay unchanged. This changes future transcription prompts, not existing transcript words.
Inputs: terms (required), guest_names (required), base_revision (required), project_path. Returns: result
set_word_automatic_toolA
Clear one word's suppression lock; suppressed waits for the next reconcile (#824).
A word set directly by set_word_suppressed_tool, an explicit-list suppress-bleed /
low-audibility apply, or Studio's Suppress toggle stays audibility_locked until this
call. It does not itself change suppressed — run reconcile_transcript_tool after
to get the computed target. Pass expected_text (the word text you read) to refuse the
unlock if the word at word_index changed meanwhile; omit to skip. TOOL_TIMEBASE: source.
source_id names the recording whose words word_index counts; it is required on a
track recorded as several takes, and the refusal is word_source_required.
Inputs: track_id (required), word_index (required), source_id, project_path, +1 more. Returns: result
set_words_ignored_toolA
Strike through and mute (or restore) a word range at render, without a cut.
Text-and-audio hide (#633): unlike suppress, the words stay in the combined
transcript; only their audio is muted, computed at render time and never
written to Clip.mute_regions. TOOL_TIMEBASE: source.
Pass expected_text (the space-joined words you read for this range) to refuse the
toggle if those words changed meanwhile; omit to skip.
source_id names the recording whose words word_index counts; it is required on a
track recorded as several takes, and the refusal is word_source_required.
Inputs: ignored (required), track_id (required), end_word_index (required), start_word_index (required), +3 more. Returns: result
set_word_suppressed_toolB
Toggle suppressed on one per-track word (TOOL_TIMEBASE: source).
Pass expected_text (the word text you read) to refuse the toggle if the word at
word_index changed meanwhile; omit to skip.
source_id names the recording whose words word_index counts; it is required on a
track recorded as several takes, and the refusal is word_source_required.
Inputs: track_id (required), suppressed (required), word_index (required), source_id, +2 more. Returns: result
set_word_timing_toolA
Set one transcript word's start and end in source seconds (Studio Adjust word timing).
word_index indexes the track's transcript, or the source_id recording's
transcript when given. Pass expected_text to refuse the change if the word was
edited meanwhile. Bounds are checked against the recording; a word that overlaps
its neighbours comes back in context.warnings. Submits
SetTranscriptWordTiming; undoable.
Inputs: end (required), start (required), track_id (required), word_index (required), +3 more. Returns: result
shorten_gaps_toolA
Shorten word gaps longer than max_gap_sec across the timeline.
Each shortened pause ripples every dialogue track; another speaker talking in a
pause needs confirmation. A ripple that cuts another track's speech stays unchanged
until you repeat with override_audio_checks=true (recorded).
Inputs: max_gap_sec, project_path, use_inaudible_opt, override_audio_checks. Returns: result
skip_session_suggestion_toolB
Skip a pending staged session (one undo step). Refuses one that is not pending.
Inputs: suggestion_id (required), project_path. Returns: result
speaker_attribute_toolC
Attribute speaker labels across the project from voiceprints (dry_run by default).
Inputs: dry_run, project_path. Returns: result
speaker_compare_pair_toolC
Compare the voices in two time windows to check they are the same speaker.
Inputs: end_a (required), end_b (required), start_a (required), start_b (required), +3 more. Returns: result
speaker_compare_window_toolC
Compare speaker identity across tracks in a time window.
Inputs: end_sec (required), start_sec (required), project_path. Returns: result
speaker_doctor_toolB
Check the speaker-identification backend is configured and working.
Inputs: no named parameters. Returns: result
speaker_enroll_toolC
Enroll voiceprints from a segment, a whole track, or every media-backed track.
Inputs: end_sec, track_id, start_sec, speaker_id, +2 more. Returns: result
speaker_gate_track_toolB
Suppress transcript words from other speakers on home tracks; use apply_transcript_gate_tool to mute audio.
Inputs: dry_run, track_id, project_path. Returns: result
speaker_label_toolC
Assign the window's speaker label on a track (dry_run by default).
Inputs: end_sec (required), track_id (required), start_sec (required), dry_run, +1 more. Returns: result
speaker_move_lane_span_toolB
Move a source span from one speaker-split lane to a sibling lane (one undo step).
TOOL_TIMEBASE: source.
Inputs: to_lane (required), from_lane (required), source_end (required), source_start (required), +1 more. Returns: result
speaker_profiles_toolB
List enrolled speaker voiceprint profiles.
Inputs: project_path. Returns: result
speaker_score_toolC
Score how strongly a time window matches the expected speaker.
Inputs: end_sec (required), track_id (required), start_sec (required), project_path. Returns: result
speaker_set_count_toolC
Set the expected number of speakers in the project.
Inputs: expected_speaker_count (required), project_path. Returns: result
speaker_split_toolA
Split a one-track recording of several speakers into a lane per speaker (dry_run by default).
Needs the speaker count (or one set with speaker_set_count_tool). enrollment maps
a speaker's name to [start, end] source-second spans of only that speaker, for some or
all speakers; the rest are clustered. Every lane plays the same media, muted where the
others talk. Flagged crosstalk plays on the likeliest speaker's lane (owner), on
every talking speaker's lane (both, twice in the mix) or on a shared lane
(lane). One undo step.
Inputs: track_id (required), names, dry_run, enrollment, +4 more. Returns: result
split_clip_toolB
Split clip(s) at a timeline timecode.
Pass track_ids for multi-track blade cuts.
When omitted, uses track_id / speaker, or all dialogue tracks.
Submits SplitAtTime on the document plane (same path as ShareCut Studio blade).
Inputs: at_time (required), speaker, track_id, track_ids, +1 more. Returns: result
stop_session_toolB
Pause DAW transport in place and clear the highlight region (GUI Pause and return (K) also returns to the play start; this tool does not).
Inputs: project_path. Returns: result
strip_silence_toolB
Rebuild a track from speech islands between detected silences.
Silence is audio under the track's own audibility floor (its voice floor) plus
over_voice_floor_db. A track whose voice does not stand out of its room has no
silence to strip, whatever over_voice_floor_db is.
use_inaudible_opt is accepted for API compatibility but ignored.
Inputs: speaker, track_id, project_path, min_duration_sec, +2 more. Returns: result
submit_document_command_toolA
Submit any document command by type: the edit a named tool does not cover.
Studio's own edits are document commands, and this reaches every type by name. The
payload is checked against the type's schema (document_command_schema_tool) and an
unknown top-level field is refused. The edit is a typed journal row and one history step,
reaches an open Studio tab and is undoable with history_undo. Prefer the named tool when one
exists: it adds the safety checks and options that tool documents. A type whose named tool
does work a bare submit cannot (InsertSegment) is refused with the tool to use. Reuse
command_id only to retry the same edit. Errors: document_command_unknown,
document_command_not_generic, invalid_payload.
Inputs: payload (required), command_type (required), command_id, project_path. Returns: result
suggest_handoff_cut_toolA
Propose retain-then-snap ripple bounds for a narrative handoff.
Omit track_id and speaker to require quiet across every dialogue lane, including muted lanes. Either selector limits analysis to one lane.
Use when cleaning up a transition, leaving a beat, or punchline-to-pivot —
not word timestamps and not preview_inaudible_cut_tool (local snap +
~0.4s absorb). Times are timeline seconds. Keeps about retain_sec
of existing air on each keep, snaps onto RMS silence islands (not
transcript gaps), and returns use_inaudible_opt: false — lock those
bounds on ripple_delete_tool.
Inputs: keep_left_end (required), keep_right_start (required), hop_ms, speaker, +4 more. Returns: result
track_addC
Add an audio track to the project from a media file.
Inputs: track_id (required), file_path (required), role, label, +2 more. Returns: result
track_add_empty_toolC
Add an empty (silent placeholder) track to the project.
Inputs: track_id (required), role, label, speaker, +1 more.
track_remove_toolC
Remove a track from the project.
Inputs: track_id (required), project_path.
track_reorder_toolC
Move a track to a new position in the track order.
Inputs: index (required), track_id (required), project_path.
track_set_media_toolC
Replace the media file backing an existing track.
Inputs: track_id (required), file_path (required), project_path.
track_set_meta_toolC
Update a track's label, role, or speaker metadata.
Inputs: track_id (required), role, label, speaker, +1 more.
track_set_mute_toolA
Mute or unmute a track in the saved mix.
Play and bounce leave a muted track out at once, and the premix after the next re-mix (Refresh). The track stays on the timeline: edits and analysis still cover it. Undoable. Solo is a per-listener GUI control, not saved.
Inputs: muted (required), track_id (required), project_path.
track_set_volume_toolA
Set a track's saved volume in dB (-60 to +12), on top of its staging gain.
The mix plays the track at gain_db + fader_db; the pipeline's balance
step never changes the fader. Undoable; stales the premix until re-mixed.
Inputs: fader_db (required), track_id (required), project_path.
transcribe_trackA
Transcribe one track, or all dialogue tracks when track_id is omitted, into word-level timestamps.
Runs even when a transcript exists, but reuses the ASR disk cache when the audio, model and prompt are unchanged (the pipeline's force_transcribe skips that cache). Replaces only those transcripts (others and extra-source transcripts are kept) and warns when it replaces a hand-edited one; the pipeline's transcribe_tracks step reuses existing transcripts instead.
Inputs: track_id, project_path. Returns: result
transcript_refine_brief_toolB
Return the transcript-refine brief: what to review before marking done.
Inputs: project_path. Returns: result
transcript_refine_done_toolB
Mark transcript refinement done, with optional notes.
Pass the status fingerprint you reviewed as expected_fingerprint to refuse if the
transcript changed since.
Inputs: notes, project_path, expected_fingerprint. Returns: result
transcript_refine_status_toolB
Report the transcript-refine gate status (blocking approve_edits while open).
Inputs: project_path. Returns: result
transcript_refine_waive_toolC
Waive transcript refinement with a recorded reason.
Inputs: reason (required), project_path. Returns: result
trim_clip_edge_toolA
Move one clip's in or out edge to source_sec (source-media seconds).
mode="ripple" moves every dialogue track by the duration change, so speakers
stay in sync: a track with a clip edge at the same instant (a session-wide cut)
moves that edge too, and every other track loses or gains the same span.
mode="gap" moves only the grabbed edge and leaves silence or fills the gap to
the neighbouring clip; nothing else moves. Expansion is clamped to unused source
between the neighbouring clips. The fix for a speech_crosses_cut hypothesis
from audition_context_tool is its evidence.fix (clip, edge, source_sec,
mode). Same operation as the DAW trim handle (TrimClipEdge); undoable. Follow
with render_preview.
The automatic audio checks hold a harsh edge or another speaker's speech and name the
finding; override_audio_checks=true moves the edge to exactly source_sec anyway
and records the override.
Inputs: edge (required), clip_id (required), source_sec (required), mode, +2 more. Returns: result
update_comment_toolB
Update comment body, tracks, and/or timeline anchor. Times are timeline seconds.
track_ids=[] makes the comment session-wide; omit it to keep the tracks.
Inputs: comment_id (required), body, track_ids, project_path, +2 more. Returns: result
update_pending_edit_toolB
Adjust the start/end bounds of a pending edit before approval.
override_audio_checks=true keeps the exact bounds given (no snapping or holds from the
automatic cut-position checks) and records the override.
Inputs: end (required), start (required), edit_id (required), snap, +2 more. Returns: result
verify_transcript_toolC
Apply bulk corrections and mark words verified on a track's transcript.
Inputs: track_id (required), corrections (required), project_path. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 218 tools
Descriptions work unusually hard to draw boundaries (e.g. "prefer set_clip_join_tool over set_join_mode_tool", "not preview_inaudible_cut_tool", "use ripple_delete_tool instead"), which rescues many pairs. However, at 218 tools there are large near-duplicate families — five `cut_*` staging tools, `ripple_delete_tool` vs `ripple_delete_text_tool` vs `delete_clips_tool` vs `propose_range_cut_tool`, `propose_edits`/`apply_edits`/`approve_edits_tool`/`apply_suggested_actions_tool`, and singular/plural pairs like `set_clip_fade_tool` vs `set_clips_fade_tool` — so misselection remains a real risk.
Most names are verb_noun (list_clips_tool, add_chapter_tool, render_preview), but the `_tool` suffix is applied inconsistently across the set: `propose_edits`, `apply_edits`, `history_undo`, `feed_read`, `record_stop_tool`, `get_transcript`, `export_transcript` coexist with `*_tool` variants of the same families. Some families also mix prefixes (history_undo vs history_status_tool, track_add vs track_add_empty_tool), which is readable but not predictable.
218 tools for a single podcast-production server is a severe mismatch against the 3-15 guidance; an agent cannot hold this surface in context or discover the right tool reliably. Even though the domain is genuinely broad, the count is well past any reasonable scoping, and several tools are marginal variants of each other.
Coverage is exhaustive for the domain: ingest/recording rooms, transcription and word-level correction, gating/bleed/audibility, cut and ripple editing, fades/joins/gain, effects, rendering and export, chapters, comments/review, review shares, feeds, segment and kit libraries, social clips, speakers, history/undo, and config. Nearly every lifecycle (create/read/update/delete/undo) has an operation, with few dead ends.
Maintenance
Related MCP Connectors
Remote MCP server for audio stem separation, denoise, reverb, and MIDI export.
An MCP server that provides tools to discover and retrieve podcast episodes transcripts.
Create, co-edit, analyze, publish, and export collaborative step-sequencer sessions through MCP.
Personal podcast queue and private RSS feed managed through a bounded MCP tool set.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceA Model Context Protocol server that exposes REAPER digital audio workstation functionality through a clean API interface, enabling programmatic control of 169+ REAPER operations across track management, MIDI editing, effects, automation and more.91MIT
- FlicenseNot gradedqualityDmaintenanceEnables AI assistants to programmatically edit, analyze, and export audio projects through MCP tools, including multi-track editing, effects, transcription, and semantic search.1-
- AlicenseBqualityDmaintenanceThis MCP server enables AI assistants to control a live REAPER DAW instance, including transport, tracks, FX, MIDI, media, markers, rendering, and project state, with an escape hatch for arbitrary ReaScript commands.40MIT
- FlicenseAqualityDmaintenanceMCP server for controlling REAPER digital audio workstation from natural language. It enables AI agents to create tracks, import audio, adjust mix parameters, and manage sessions through typed operations and a REAPER bridge.22-
social_clip_report_toolB
Return a Markdown listing of social clip candidates.
Inputs: project_path. Returns: result