RELATED 01
ffmpeg-skill
Local video and audio editing through 42 FFmpeg-backed tools, as an Agent Skill and a stdio MCP server
4.7kGitHub stars
9.1k/wknpm downloads
86popularity score
Kinocut
Local FFmpeg-based video editing with typed tools, preflight checks and hashed receipts
By KyaniteLabs Python Remote (hosted) · Streamable HTTP Media
64 popularity score ★ 253 ↓ 8.9k/mo
Connect your client
Remote (hosted)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Its docs show no set-up for a particular client: add this address in your client the way it adds a remote MCP server.
Addressits hosted endpoint
https://mcp.kyanitelabs.tech/mcpSettings 6
| Set together: tool surface | |
|---|---|
KINOCUT_TOOLS |
Optional. |
| Set together: hyperframes | |
KINOCUT_HYPERFRAMES_TELEMETRY |
Optional. |
MCP_VIDEO_HYPERFRAMES_COMMAND |
Optional. |
| Set together: revideo | |
KINOCUT_REVIDEO_EXECUTABLE_PATH |
Optional. |
| Set together: vision | |
KINOCUT_VISION_MODE |
Optional. |
KINOCUT_VISION_BASE_URL |
Optional. |
What it does
Kinocut is a video editing server from KyaniteLabs that runs on your own machine and puts typed tools on top of FFmpeg. It used to be called mcp-video, and it also ships a Python client and a `kino` command line. An assistant calls named operations with checked parameters instead of writing raw FFmpeg flags. These include trimming, merging, cropping, resizing, converting, burning in subtitles, adding text and watermarks, normalising loudness and splitting HLS segments. Many operations write to a staged file first and only replace the destination once the render has passed its checks.
On top of the basic edits it has larger pipelines. A workflow engine takes a JSON job spec built from an allowlist of operations (probe, trim, resize, convert, crop, merge, add_text, composite_layers, burn_in), dry-runs it, renders it and writes a receipt with input and output hashes for each step. There are also repurposing tools that turn one recording into vertical, horizontal and square packages, a release checkpoint that acts as a quality gate before publishing, a review-first rescue pipeline, 360 equirect assembly, and review tools that cannot approve media without a stored human decision. Optional extras add Whisper transcription, stem separation, AI upscaling, image analysis and HTML-based video through Hyperframes.
Version 1.16.1 has 203 MCP tools. The source sets hard caps that the README does not list: input video up to 4 hours, 8K resolution and 4 GB, batches of up to 50 files, a 10-minute default FFmpeg timeout, and transcription of audio up to 1 hour (kinocut/limits.py). Unless you give an output path, results are written next to the input as `<name>_edited.<ext>`, and the name is changed so the input is not overwritten (kinocut/paths.py).
What it needs
| To run it yourself | Python 3.11+ |
|---|---|
| Also |
|
Python 3.11 or newer, and uv for the `uvx` launch command (or `pip install kinocut`). FFmpeg and ffprobe must be on PATH; `video_stabilize` needs an FFmpeg build with vidstab. Runs on macOS, Linux and Windows. Optional features need pip extras such as `kinocut[transcribe]`, `[stems]`, `[upscale]`, `[image]`, `[audio]` or `[ai]`, some of them several GB. Hyperframes tools need Node.js 22+ and the Hyperframes CLI. Revideo tools need Node.js 18+, npm and a Chromium-family browser. `kino doctor` reports what is missing.
Limits
Caps in its source: video up to 4 hours, 8K and 4 GB per input; batches of up to 50 files; transcription up to 1 hour of audio; FFmpeg runs time out after 10 minutes by default. It does not stitch raw 360 files (`.insv`, GoPro `.360`), so export from the camera app first. Layered compositing outputs video only, and some blend geometries are refused with `unsupported_blend_geometry`. Its makers say caption speech does not translate text, no generative adapter exists, and the Claude Desktop MCPB package is unsigned, staged and not a sandbox. The registry lists a hosted endpoint, but the README describes no hosted service, only local use.
Who it suits
People who want an agent to cut, caption and check local video files on their own machine, from short vertical clips to longer edits that need a review step.
Tools 20 of 205
8 only read, 12 change things.
video_info- Read a file's duration, streams, resolution and codecs.
video_trimchanges- Cut a clip to a start time and duration.
video_mergechanges- Join several clips into one file.
video_add_textchanges- Burn one text overlay into a video.
video_add_textschanges- Burn several timed text overlays in one pass.
video_add_audiochanges- Add or replace a video's audio track.
video_resizechanges- Change resolution or aspect ratio, for example to 9:16.
video_convertchanges- Convert a video to another container or codec.
video_speedchanges- Speed a clip up or slow it down.
search_tools- Find hidden tools by keyword without registering them.
video_extract_framechanges- Save a single frame as an image.
video_validate_text_layout- Check planned text overlays for overflow and placement problems.
video_quality_check- Report on visual and audio quality of a finished file.
video_review_packagechanges- Gather frames and evidence for a human review.
video_workflow_inspect- Inspect the state of a multi-step workflow job.
video_rescue_inspect- Diagnose a damaged or problem file before a repair.
hyperframes_doctor- Check that Node and the Hyperframes CLI are ready.
video_doctor- Check the local setup, such as FFmpeg, without changing anything.
video_shortschanges- Trim chosen ranges, caption them and export one 9:16 Short with a receipt.
video_project_createchanges- Create a project scaffold for a cinematic edit.