Talk-n-Edit
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Talk-n-Edit is a prompt-driven video editor: Claude designs the edit as a timeline (JSON) and the server renders it. Start by asking the user for an editing brief (style, pacing, captions, music, effects) unless they gave one; catalog_list styles has presets. Workflow: project_open -> media_import -> analyze (transcript, faces, scenes, beats) -> transcript_get / analysis_get -> timeline (catalog_list: components, transitions, looks, sfx, styles) -> timeline_validate -> render_start draft -> frames_get / open_in_player -> revise with timeline_get + timeline_validate -> render_start final -> qa_check. For effects the library lacks, write one with fx_create and show its preview first. Jobs return a jobId: job_status can wait up to 45 s. Finals are copied to the Clips folder (system_doctor shows the Inbox and Clips folders). Never invent ids, components or presets.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| system_doctorA | Check ffmpeg, the Python analysis environment, fonts and folders. Run this first if anything fails. |
| project_openB | Open a project by name (creates it if missing) and list its assets, analyses, timelines and renders. With no name, lists projects. |
| media_importA | Import video, audio or image files (or every media file in a folder) into a project. Copies originals into the workspace and builds a 540p proxy. Returns a jobId. |
| media_inspectB | Probe an asset (duration, size, fps, audio, analyses done) and return a 4x3 contact sheet image of evenly spaced frames. |
| analyzeA | Analyse an asset. transcript = word-level speech-to-text (clip picking, captions, silence removal); faces = face positions (auto-face / split reframing); scenes = shot boundaries; beats = music tempo + beat times (beat-synced cuts). Cached; returns a jobId. |
| transcript_getA | Read an asset's transcript as numbered sentences with times: [s0012 03:41.2–03:47.9] text. Use sentence ids in clips ({sentences:{from,to}}) or the times as in/out. format 'words' gives per-word times. |
| frames_getA | Grab up to 12 frames (as images) from a video asset or a finished render at the given times in seconds. Use it to check framing, hooks and cuts. |
| catalog_listB | List what timelines can use: components (text overlays + animated graphics + custom fx, with prop schemas), styles (saved edit styles), transitions, looks, sfx, music, luts, captionPresets, brandKits, fonts. |
| timeline_validateB | Validate a timeline JSON and save it as a new version. Returns resolved duration per clip, the spoken text each clip will contain, warnings, and errors with JSON paths. Same name + project continues the same timeline id. |
| render_startA | Render a timeline to MP4. quality 'draft' = fast 540-wide preview; 'final' = full quality, loudness-normalised, also copied to the export folder. Give timelineId (+version) or an inline timeline. Returns renderId + jobId. |
| job_statusA | Status of one or more jobs (import, analyze, render, qa). Waits up to |
| job_cancelC | Cancel a queued or running job. |
| qa_checkA | Check a finished render: format, duration, black frames, dead air, loudness, hook timing; returns a contact sheet and frames after the first cuts. Returns a jobId. |
| analysis_getA | Read analysis results: scenes (each shot with start/end, face coverage, face position and size — use it to pick shots), beats (tempo + beat times for beat-synced cuts), faces (same per-scene view). |
| timeline_getA | Read a saved timeline (latest or a given version) so you can revise it: change what the user asked, then timeline_validate saves the next version. |
| style_saveB | Save the user's editing preferences as a named style (look, transition, autoPunch, captions, sfxOnTransitions, hook, removeSilences, music volume, notes). Use it in timelines as "style": "" so every future edit matches. |
| fx_createA | Create or update a custom animated effect when the built-in components can't do what the user wants. Write a React component (TSX) for Remotion:
|
| open_in_playerA | Open a finished render in the default video player (QuickTime on a Mac) so the user can watch it. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 18 tools
Most tools target a distinct resource+action and are well separated (import vs inspect vs analyze vs render vs qa). The main near-overlaps are analyze vs analysis_get and frames_get vs media_inspect/analysis_get frame output, though the descriptions clarify trigger-vs-read and the differing frame purposes.
The dominant pattern is resource_verb (media_import, timeline_validate, render_start, job_status, style_save, fx_create), used consistently across most of the surface. Deviations like the bare verb 'analyze', 'system_doctor', and 'open_in_player' are minor and still readable.
18 tools is slightly heavy but each maps to a real stage of the video-editing pipeline (setup, ingest, analysis, authoring, render, QA, delivery). No tool appears clearly redundant, though the ceiling is near the upper bound for comfortable scoping.
The lifecycle is well covered end to end: doctor, project open, import, analyze, transcript/frames read, catalog, timeline validate/get, render, job status/cancel, QA, style save, fx create, and player open. Minor gaps exist around deletion/cleanup of projects and assets and listing renders, which agents can mostly work around.