Kinocut

Your agent edits the video. You approve the cut.

Kinocut is a local, guardrailed video-editing MCP server, Python client and kino CLI built on FFmpeg. Every edit is checked before FFmpeg runs, writes a new file unless you allow an overwrite, and returns a structured result. Quality gates catch black frames, silence and bad loudness, and a Video Receipt records what ran. Formerly mcp-video.

What your agent sees

By default the MCP server lists 18 tools: an 8-tool hot path (video_info, video_trim, video_merge, video_add_text, video_add_audio, video_extract_frame, video_shorts, video_doctor), eight operation families that reach every other operation, and list_operations / describe_operation to find and fetch them. KINOCUT_TOOLS=profile:<name> lists one job's tools by name and KINOCUT_TOOLS=all lists all 206.

video_review_card turns a finished cut into a review card with a contact sheet and Approve, Dig in and Try another range buttons for Hermes Desktop chat.

Install and inspect

Requires Python 3.11 or newer, plus FFmpeg and ffprobe on PATH.

pip install kinocut==2.0.0
kino doctor
kino info input.mp4

Agent plugins launch a specific published version through uv. Follow the plugin instructions for the host you use. Kinocut also provides from kinocut import Client; the older mcp_video import remains supported.

Output safety

Prefer new output paths. Replacing an existing output requires explicit overwrite=true (CLI --overwrite). Existing non-media artifacts also require evidence that the same durable job created them. Agent configuration homes are protected. Use one writer per output, check the result, and require human visual/audio review before publishing.

Optional model calls and network use

Ordinary local FFmpeg editing and color grading need no account, model SDK or API key. Kinocut's optional vision calls default to KINOCUT_VISION_MODE=off. For observations with your chosen provider, prepare local images through Kinocut and inspect them with the host agent's image tools, including local models. Kinocut does not automatically inherit the host's model configuration.

Kinocut's own optional product-description adapter requires an Anthropic-compatible endpoint and explicit local or remote configuration. The separate sampled-frame Anthropic QC path also requires explicit remote permission and configuration. A local/custom product route cannot activate it.

Installing packages downloads dependencies. URL inputs download requested media; font and model-weight features may fetch assets from GitHub. Explicit Revideo installation runs npm ci and may download renderer/browser dependencies. Revideo telemetry defaults off. Hyperframes is a separately installed CLI: Kinocut sets telemetry opt-outs, skips passive update checks and disables implicit capture/snapshot model analysis. Docker rendering and browser previews without a verified opt-out require explicit telemetry consent.

These controls are not a network sandbox. Inspect executable scenes and project assets before use. See the vision and network policy for configuration, supported preview versions and limits.

Documentation · Release notes · Issues · Website