fcpxml-mcp-server
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| OPENAI_BASE_URL | No | Route LLM calls through any OpenAI-compatible proxy (LiteLLM, OpenRouter, Ollama, vLLM) | |
| FCP_PROJECTS_DIR | No | Root directory for FCPXML file discovery via list_projects | ~/Movies |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| inspectA | Read a timeline or project without changing it. Use this first to understand what you are working with. Actions: list_projects, analyze_timeline, analyze_pacing, list_clips, list_markers, list_roles, list_keywords, list_effects, list_templates, list_library_clips, list_compound_clips, list_connected_clips, filter_by_role. |
| diagnoseA | Find problems in a timeline: gaps, flash frames, duplicates, dead air, and beat structure. Read-only. Run before editing. Actions: validate_timeline, detect_gaps, detect_duplicates, detect_flash_frames, detect_silence_candidates, detect_media_silence, detect_beats, find_short_cuts, find_long_clips, diff_timelines. |
| editA | Change clips on the timeline: insert, delete, trim, split, reorder, retime, remove silence, and attach audio, B-roll or transitions. Writes a new file. Actions: insert_clip, delete_clips, trim_clip, split_clip, reorder_clips, change_speed, rapid_trim, add_transition, add_audio, add_connected_clip, assign_role, fill_gaps, fix_flash_frames, remove_silence_candidates, remove_media_silence. |
| markB | Add or import markers and chapters, including from SRT/VTT subtitles, transcripts, and beat analysis. Actions: add_marker, batch_add_markers, import_srt_markers, import_transcript_markers, import_beat_markers, snap_to_beats. |
| generateB | Build new timeline structure from source clips: rough cuts, montages, A/B roll, templates and compound clips. Actions: auto_rough_cut, generate_ab_roll, generate_montage, apply_template, create_compound_clip, flatten_compound_clip, import_edl_json. |
| transcriptB | Transcribe source media locally and edit the timeline by what was SAID rather than by timecode. Also removes filler words. Actions: transcribe_media, edit_by_transcript, remove_filler_words, transcript_pack. |
| deliverA | Get the edit out: export to other NLEs, CSV, EDL and stems, reformat, relink media, or push straight into a running Final Cut Pro. Exports and push refuse a cut with no rendered preview of its current state (pass confirm_unreviewed=true to ship anyway). Actions: export_csv, export_edl, export_fcp7_xml, export_resolve_xml, export_role_stems, reformat_timeline, relink_media, push_to_fcp, list_fcp_libraries. |
| previewA | See the edit without opening Final Cut Pro. Render a proxy video of the timeline, a contact sheet of every cut, a single frame, or a filmstrip-plus-waveform check read from the SOURCE MEDIA rather than from the XML. Run preview_check to confirm a fix actually landed. Actions: preview_render, preview_sheet, preview_frame, preview_check, preview_timeline. |
| watchA | Close the round-trip. Watch the Final Cut Pro export folder, detect the XML the moment it lands, and diff it against the last one seen. Pair with deliver.push_to_fcp for a full loop: push in, edit, Cmd-E, watch_pull. Actions: watch_start, watch_status, watch_stop, watch_pull. |
| indexA | The analysis cache under ~/.fcp-mcp/index.db. index_status shows what is cached and how old it is; index_build warms it for every source in a timeline (args: filepath, with_transcript); index_clear drops it. Every other tool works with the index off (FCP_MCP_INDEX=off) — this only makes the second question fast. Actions: index_status, index_build, index_clear. |
| scenesA | Shot boundaries. detect_scenes reads every source in a timeline and reports cuts in timeline time (args: filepath, clip_name?, backend auto|content|adaptive|ffmpeg, threshold?, min_scene_len=0.5); scenes_to_markers writes a marker at each cut; scenes_split cuts the clips there. PySceneDetect via the scenes extra, ffmpeg without it — the result names which one answered. Actions: detect_scenes, scenes_to_markers, scenes_split. |
| organizeA | Bulk library logging and the ledger. organize_auto proposes keywords per clip from cached captions and the transcript (apply=true writes them). Select clips with clip_name (glob), keyword and/or role, then organize_keywords (keywords, mode add|remove|replace), organize_rate (rating favorite|rejected|clear) or organize_roles (audio_role, video_role); each writes a _organized copy. history lists every recorded operation for the file's folder (limit); undo (n) moves the outputs of the last n writes into the journal's undone/ folder — it never deletes and refuses when a file changed since it was written. Actions: organize_auto, organize_keywords, organize_rate, organize_roles, history, undo. |
| findA | Natural-language shot search. find_shots (filepath, query, limit=10, visual=false, clip_name?) ranks moments by what was said (transcript), what was logged (names, keywords, notes, markers, audio events) and, when a local vision model is installed, what the frames look like — the first line of every result names which tiers answered and why one could not. find_index (captions=true, backend) warms scenes and captions and reports which clips have no transcript (it never transcribes or goes online). find_to_timeline (min_source_separation=1, output_path?) assembles the hits into a _found selects reel and reports its diversity score. Actions: find_index, find_shots, find_to_timeline. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
| qc-check | Run a full quality control check on your timeline — flash frames, gaps, duplicates, and health score |
| youtube-chapters | Extract chapter markers formatted for YouTube descriptions |
| rough-cut | Guided rough cut generation — choose keywords, pacing, and duration |
| timeline-summary | Quick overview of a timeline — stats, pacing, and potential issues |
| cleanup | Find and fix common timeline issues — flash frames, gaps, and duplicates |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 13 tools
The thirteen tools map onto distinct workflow stages—inspect, diagnose, edit, deliver, preview, watch, etc.—so an agent can usually pick the right one by intent. Minor overlap exists between inspect and diagnose (both read-only analysis) and between scenes_split and edit.split_clip, but the descriptions separate general understanding from problem-finding and scene-boundary work.
All tool names follow a consistent single-lowercase-word style, which is predictable after seeing a few. The main deviation is grammatical: most are verbs (inspect, edit, mark, watch, find) while transcript, index, and scenes are nouns, and the actual operations live inside as snake_case actions.
With 13 tools, the server sits comfortably in the well-scoped range, and each tool represents a distinct phase of FCP XML editing: analysis, diagnostics, editing, marking, generation, transcription, delivery, preview, round-trip, indexing, scene detection, organization, and search. None feels redundant; the breadth is justified by the size of the domain.
The surface is unusually complete for the domain, covering inspection, diagnosis, editing, generation, metadata logging, transcription, export, preview, watch-loop integration, caching, scene detection, organization, and natural-language search. Read-only tools have write counterparts, output paths are verifiable, and there are no obvious dead-end workflows.