Final Cut Pro editing via FCPXML

SkillFiles & storage

Edit Final Cut Pro timelines with natural language via FCPXML. Use when the user wants to analyze, cut, trim, mark up, transcribe, montage, or export a Final Cut Pro project, mentions .fcpxml or .fcpxmld files, asks to remove silence or filler words from a video edit, wants a rough cut assembled from source clips, wants markers or chapters added from a transcript or SRT, or wants an edit pushed into a running Final Cut Pro. Requires the fcp-mcp-server MCP server.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Final Cut Pro editing via FCPXML skill

What this skill tells your AI

The instructions your AI receives, as published by daredev256/fcp-mcp-server in skill/SKILL.md and read by ahel’s review.

Before you touch anything

Always inspect then diagnose before you edit. Timelines carry structure that is not obvious from a filename, and an edit against a wrong assumption is expensive to undo.

  1. inspect with action analyze_timeline for shape, duration, resolution.
  2. diagnose with action validate_timeline for gaps, flash frames, duplicates.
  3. Read the preview://<path> resource to actually SEE the timeline before you change anything. It renders a self-contained HTML timeline, including connected clips on their own lane rows, so you can see B-roll and audio layers, not just the primary storyline.

The seven groups

GroupUse it for
inspectRead-only understanding. Start here, always.
diagnoseFind problems: gaps, flash frames, dead air, duplicates, beats.
editChange clips: insert, delete, trim, split, reorder, retime, remove silence.
markMarkers and chapters, including SRT/VTT and beat import.
generateBuild new structure: rough cuts, montages, A/B roll, templates, and import_edl_json from a video-use cut list.
transcriptTranscribe locally, then edit or clean up by what was SAID. transcript_pack puts the whole shoot on one page — read it before planning a dialogue edit. Every action takes backend: "elevenlabs" for speaker labels and audio events; that uploads the media to ElevenLabs and needs ELEVENLABS_API_KEY — the default local never leaves the machine.
deliverExport to other NLEs, reformat, relink, push into FCP.
previewSEE the edit: proxy render, contact sheet, and a filmstrip+waveform read from the source media.
watchClose the round-trip: notice the operator's Cmd-E export and diff it against the last one.
indexThe analysis cache. index_status says how old it is; index_build warms every source in a timeline; index_clear drops it. Nothing depends on it.
scenesShot boundaries from the pixels. detect_scenes lists cuts per clip in source and timeline time (PySceneDetect when installed via [scenes], else ffmpeg's coarser scene filter); scenes_to_markers drops a marker on each; scenes_split cuts the clips there. Results are cached in the index.
organizeBulk library logging and the ledger. organize_auto proposes keywords per clip from cached captions and the transcript (apply=true writes them). Select clips by clip_name glob, keyword and/or role, then organize_keywords (add/remove/replace), organize_rate (favorite/rejected/clear) or organize_roles — each writes a _organized copy. history lists every recorded operation for the file's folder; undo moves the last n outputs into the journal's undone/ folder (never deletes, refuses if a file changed since).
find"Where is the bit where…" as a router. find_shots ranks moments by what was said (transcript), what was logged (names, keywords, notes, markers, audio events) and, with the [find] extra and a downloaded local model, what the frames look like — the first line names which tiers answered and why one could not. find_index warms scenes and captions and reports which clips have no transcript (never transcribes, never online). find_to_timeline assembles the hits into a _found selects reel with a diversity score.

Every call takes {"action": "...", "args": {...}}. If you pass an action the group does not own, the error lists the valid ones — read it rather than guessing.

FCPXML facts that will bite you

  • Time is rational, never float. 3600/2400s is 1.5 seconds. Never do float math on timecode and never round it yourself.
  • offset is the timeline position. start is the source in-point. They are different numbers and confusing them silently shifts an edit.
  • Library clips and timeline clips are different elements. <asset-clip> lives in the library, <clip> lives on the timeline.
  • Markers are children of clips, not siblings.
  • The <spine> is the primary storyline. Connected clips hang off spine clips by lane: positive is above (video), negative is below (audio).
  • .fcpxmld bundles are directories, wrapping Info.fcpxml plus sidecar data. Sidecars must be preserved on save or object-tracking and Cinematic data is destroyed.
  • Duplicate clip names are common. Prefer element-based operations over name lookups where a tool offers both.

Silence removal: pick the right one

There are two detect-then-fix pairs, and they are not interchangeable. Both follow the same split as every other detect/fix pair in this server: the detector lives in diagnose, the fixer lives in edit.

  • detect_media_silence (diagnose) → remove_media_silence (edit) reads the actual source audio with ffmpeg. Accurate, slower, needs ffmpeg installed.
  • detect_silence_candidates (diagnose) → remove_silence_candidates (edit) uses timeline heuristics only. Fast, no ffmpeg, and it guesses.

Use the media versions when correctness matters. Say which one you used.

Working order that tends to hold

  1. inspectdiagnose → read the preview:// resource.
  2. transcript with action transcript_pack if the edit is dialogue-driven — one page of everything said, so the plan is made from the words rather than the clip names.
  3. generate for the assembly, edit for the refinement, mark for chapters.
  4. preview with action preview_check over the range you changed, before offering the edit as done. This is what makes the edit non-blind. The preview:// resource and preview_timeline both draw from the XML — they show what was WRITTEN, so they cannot tell a fixed cut from a broken one. preview_check reads the media. Do not substitute one for the other, and do not skip it because the tool call reported success.
  5. deliver, either exporting or push_to_fcp into the running app.
  6. watch with action watch_start once per session, then watch_pull after the operator exports. Final Cut Pro has NO programmatic export — Apple never shipped one — so the loop only closes when they press Cmd-E. Say so plainly rather than waiting in silence.

Do not

  • Do not edit without a diagnose pass. Flash frames and gaps compound.
  • Do not assume ffmpeg or Whisper are installed. Both degrade gracefully and return nothing rather than erroring, so check the response.
  • Do not read a preview_render summary past the word UNVERIFIED. It means the rendered file's duration could not be read back, so nothing about that render has been confirmed.
  • Do not ignore a Substituted: line. The proxy renders transitions as hard cuts, so a dissolve in the timeline is NOT what you are looking at.
  • Do not claim an edit landed in Final Cut Pro unless push_to_fcp was called and reported success. Writing an .fcpxml file is not the same as importing it.
  • Do not trust a tool's own success message as proof the timeline looks right. Read preview:// and look.

Signals

GitHub stars
105
Forks
19
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
final-cut-pro
Source
github.com/daredev256/fcp-mcp-server