mcp.film / Editing & Post / claude-real-video (crv)
claude-real-video (crv)
Lets an agent actually watch footage locally — scene-cut keyframes plus transcript, no upload.
What it does
crv turns a local file or a URL into something an LLM can read: it detects scene changes with ffmpeg rather than sampling on a fixed interval, deduplicates near-identical frames by pixel difference, transcribes the audio with Whisper, and hands back frames plus a manifest. The MCP mode adds a memory layer — every watched video is indexed, so an agent can ask questions across the whole library instead of one clip at a time. For a filmmaker-agent this is the dailies-review bench that TwelveLabs and Valossa do in the cloud, done on your own machine with nothing leaving it.
Connect
claude mcp add crv -- crv-mcp
Local install: crv-mcp (after `pip install 'claude-real-video[mcp]'`)
Sample tools
watch_videoget_framesget_transcriptsearch_memorylist_watchedField notes
MIT, 2,250 stars and 202 forks off a repo created 2026-06-30, last pushed 2026-10-08; PyPI is at v0.10.7 (2026-09-26), and the server was registered as io.github.HUANGCHIHHUNGLeo/claude-real-video. The main trap: MCP is an extra, not the default. Plain `pip install claude-real-video` gives you the `crv` CLI with no server, and the `claude mcp add` line above only works after installing the `[mcp]` extra. The old branch confusion is fixed — `master` is now the default branch and its README documents the MCP mode, the five tools and `crv-mcp` directly; a stale `main` branch still exists, so ignore it if a link lands you there. The README says MCP mode is verified end-to-end on Claude Code only; Claude Desktop and Cursor speak the same stdio protocol but are untested. FFmpeg must be on PATH and is not pip-installable; Python >=3.10; transcription needs the `[whisper]` or `[fast]` extra, which is a large download (`[mlx]` and `[speakers]` extras also exist). Analyses cache under ~/.cache/crv-mcp, so repeat questions are cheap but the directory grows without bound. Frames come back as images, so a long video is expensive in tokens — reach for get_transcript when you only need the words. Free and local, but the upsell has grown: 'crv Pro' is now a one-time $29 (was $19) adding camera, sound and emotion analysis, sold via Capafy and Lemon Squeezy, plus a separate $9.90 'Real Video for Codex' add-on. Downloading from a URL is your legal problem, not the tool's.
Also in Editing & Post
Hosted video-understanding MCP: transcripts, shots, visual scenes, soundbites, and clips.
A local editor with no GUI and no LLM: an append-only edit graph an agent drives over MCP.
Local MCP and CLI toolkit for agent-driven video composition, editing, analysis, and rendering.
Hosted MCP that clips, captions, dubs, reframes, and publishes long-form footage.
Deterministic video from one JSON template — same input, same frames, rendered locally.
Guardrailed local video editing MCP — FFmpeg, receipts, quality gates, and repurposing.
Pairs well with
Quoted dubbing, subtitles, transcription and multi-voice audiobooks from one remote MCP; a human approves each price.
Official Deepgram server: Nova transcription, diarization, and audio intelligence.
Official hosted Airtable MCP: query, create, and update production-tracking bases.