footage-mcp

This server is the footage workbench's agent door — the step before either video engine, where a raw recording (a talking-head take, a screen capture, an explainer) becomes a clean, condensed cut. Its 15 footage_* tools cover the whole preparation loop: register source clips, transcribe them, propose cuts, validate and refine the edit list, trim and split, apply a color grade, inspect the timeline, and render the finished cut. A thin front over the automation daemon, like its creation-vertical siblings.

The working object is the edit decision list: footage_propose_cuts reads the transcript and proposes what to remove — fillers, retakes, dead silence — and footage_validate_edl checks any edit list before it renders. That keeps the judgment step (what stays in the take) reviewable as plain data rather than buried in a render, and it is why the prepared output — the clean cut plus its transcript and edit list — drops straight into either engine's brief as an asset.

How your agents use it

• "Tighten this take." — footage_add_source, footage_transcribe, footage_propose_cuts, review, then footage_render for the clean cut.
• "Show me what the cut looks like before rendering." — footage_timeline_view lays out kept and removed segments; footage_trim_clip and footage_split_clip handle manual adjustments.
• "Match the house look." — footage_list_grades lists the available color grades to apply at render.

Prerequisites

The automation daemon must be running (Dashboard ▸ Platform Health); transcription and rendering run locally on your machine. Daemon settings live in the configuration reference.

Tool reference

Tool	Parameters	What it does
footage_health	—	Footage-prep health: ffmpeg/ffprobe present, Grok-STT availability, render concurrency. No args.
footage_list_grades	—	List available color-grade presets (subtle, neutral_punch, warm_cinematic, none) plus 'auto' (data-driven). No args.
footage_transcribe	video_paths*: array, brief_id: string, revision_id: integer, engine: string, provider_name: string, transcript_source: string, language: string, diarize: boolean, keyterms: array, filler_words: boolean	Transcribe source mp4(s) to a word-level transcript.json + captions.json (HF caption shape). Uses Grok STT by default (word-level + diarization); pass transcript_source to ingest an externally-supplied SRT/VTT/word-JSON instead. Async — returns a job_id; poll with footage_get_job. Outputs land in the brief's assets/ (when brief_id given) or a standalone store.
footage_propose_cuts	transcript_path*: string, source_name*: string, source_path*: string, silence_threshold: number, pad: number, grade: string	Produce a MECHANICAL baseline EDL that removes silences ≥ threshold from a transcript (deterministic, no judgment). Returns an EDL object to refine: drop retakes, pick the best take, tighten timing — that is the agent's editorial work. Single source.
footage_validate_edl	edl*: object, base_dir: string, check_files: boolean	Validate an EDL object before rendering: sources non-empty, every range.source resolves, end>start, files exist. Returns {valid, errors[], expected_duration_s}.
footage_render	edl: object, edl_path: string, brief_id: string, revision_id: integer, engine: string, output_dir: string, output_name: string, quality: string, draft: boolean, build_subtitles: boolean, transcript_paths: object, edl_base_dir: string	Render an EDL → cut.mp4 via the full pipeline (per-segment extract with 30ms fades + grade + HDR tonemap → lossless concat → overlay composite → subtitles burned LAST → two-pass loudnorm to -14 LUFS). Async — returns a job_id; poll with footage_get_job. Provide either `edl` (object) or `edl_path`. Output lands in the brief's assets/ (when brief_id given) as cut.mp4 + edl.json.
footage_get_job	job_id*: string	Poll one footage-prep job (render or transcribe) by job_id. Returns status (queued|running|done|failed), outputs, and any error.
footage_list_jobs	limit: integer	List recent footage-prep jobs (most recent first).
footage_list_sources	brief_id*: string, revision_id: integer, engine: string	List a project's raw footage — the videos in the brief's assets/raw/ (+ any transcript/captions sidecars). This IS the project footage both the operator UI and agents share. Requires brief context.
footage_add_source	brief_id*: string, source_path*: string, revision_id: integer, engine: string	Import a raw video into the project (copies it into the brief's assets/raw/) so the project is self-contained and the footage is accessible to both operator and agents. Returns the registered source. Idempotent.
footage_remove_source	brief_id*: string, name*: string, revision_id: integer, engine: string	Remove a raw source (+ its transcript/captions sidecars) from a project's assets/raw/.
footage_pack_transcripts	brief_id*: string, revision_id: integer, engine: string	Layer 1 (the primary reading view) — pack ALL the project's transcribed footage (assets/raw/*.transcript.json) into ONE phrase-level takes_packed.md: `[start-end] S<spk> text` per phrase, ~1/10 the tokens of raw word JSON. Transcribe sources first. Returns {path, markdown, sources}. Read this to reason about cuts efficiently.
footage_timeline_view	video_path*: string, start*: number, end*: number, n_frames: integer, brief_id: string, revision_id: integer, engine: string	Layer 2 (on-demand visual) — generate a filmstrip + waveform PNG for a time range of a clip, so you can SEE the footage at a decision point (ambiguous pause, cut-point sanity, or self-eval of a rendered output at a cut boundary). Returns {image_path}; then use the Read tool on that PNG to inspect it. Pure ffmpeg.
footage_trim_clip	brief_id*: string, source_path*: string, start*: number, end: number, out_name: string, revision_id: integer, engine: string	Non-destructively trim a clip to [start,end] → a NEW clip in the project (assets/raw/); the source is untouched. Frame-accurate, resolution preserved (a plain cut — no grade/fades). Returns {path,name,start,end}.
footage_split_clip	brief_id*: string, source_path*: string, split_points*: array, revision_id: integer, engine: string	Non-destructively split a clip at point(s) → N NEW segment clips in the project (assets/raw/); the source is untouched. e.g. split_points=[12.5] → _part1 (0–12.5) + _part2 (12.5–end). Returns {parts:[{path,name,start,end}]}.

Where to go next

• Video — the production studio — the footage workbench in the app, and the engines the output feeds
• remotion-mcp · hyperframes-mcp — the two engines a prepared cut lands in