Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Kinocut

Local FFmpeg-based video editing server with guardrails: trims, captions, shorts and quality gates, 18 tools by default and 206 in all.

Reviewed by ReviewRamble editors · Updated

Setup docs Source codeOpen source

Our rating

3.9/5

A video server built around not wrecking your footage, with the deepest tool catalogue in its niche and a release cadence that asks you to pin versions.

What is Kinocut?

Kinocut, from KyaniteLabs and formerly called mcp-video, is a local video-editing MCP server, Python client and kino command-line tool built on FFmpeg. An agent can trim, merge, caption, resize, crop, overlay, mix audio and cut vertical shorts, then run quality gates for black frames, silence and loudness before a person sees the result. Each job can produce a Video Receipt recording what was asked, what ran and what changed.

It is Apache-2.0, runs on macOS, Linux and Windows with Python 3.11+ and FFmpeg, and needs no account. Version 2.0.1 shipped on 10 November 2026. The repo has about 264 stars and 59 forks, and the kinocut package had 8,854 PyPI downloads in the last month, plus 3,024 for the old mcp-video name, which now installs Kinocut.

How to install Kinocut

uvx --from kinocut==2.0.1 kino --mcp

Is Kinocut worth it?

Kinocut's design starts from what goes wrong when an agent drives FFmpeg directly: guessed flags, overwritten sources and outputs nobody looked at. Every edit is validated before FFmpeg runs, writes to a new file unless overwrite is set, and returns a structured result. The security policy adds that relative output paths with .. traversal, symlinks, system directories and sensitive dotfiles are rejected, and an output that resolves to one of the same call's inputs fails with invalid_output_path. Absolute paths are allowed by design, so an agent can still write anywhere the user can.

The tool list is handled better than most large servers. By default a client sees 18 tools, about 22 KB of JSON: eight hot-path tools such as video_trim, video_add_text and video_shorts, eight operation families, and two discovery calls that fetch any other operation's schema on demand. KINOCUT_TOOLS=all lists all 206 tools at about 178 KB, and 19 named profiles sit in between.

Captions are less turnkey than the tagline suggests. Burning in an SRT works from the core install, but generating one needs the transcribe extra, which pulls in openai-whisper and torch and runs locally. Optional vision features are off by default; in remote mode the product adapter and the QC worker send frames to api.anthropic.com with your key. Kinocut publishes nothing itself, and the README's own hosting suggestion is a third-party community stack.

The catch: this is a fast-moving project. Versions 2.0.0 and 2.0.1 both shipped on 10 November 2026, the install snippet pins kinocut==2.0.1, and the docs folder runs to dozens of release, gate and governance files. One open issue reports that Approve & export can ship a cut that failed QC, which matters because the review card is the main safety step. Media is also untrusted input: a transcript or a captured web page from the Hyperframes tools can carry text an agent might follow.

Where it earns its place

  • A solo podcaster publishing weekly

    Turn a 60-minute episode into three captioned vertical shorts

    Install the transcribe extra, have the agent transcribe locally, pick ranges and call video_shorts, then run the loudness gate. No per-minute fees; the cost is model tokens and a local Whisper run.

  • A developer advocate at a SaaS company

    Trim and caption product demo recordings for the docs site

    Trim with video_trim, burn in an SRT with video_add_text or subtitles, resize to 16:9, and keep the Video Receipt with the asset so reviewers can see exactly what was cut.

  • A small studio reviewing AI-generated clips

    Find unusable takes before an editor spends time on them

    Run the quality-control operations for black frames and silence across a folder, then use video_review_card or kino release-checkpoint for a contact sheet and pass/fail report.

Before you commit

Setup
Install FFmpeg, then claude mcp add kinocut -- uvx --from kinocut==2.0.1 kino --mcp, or the same uvx command in a JSON config for Claude Desktop, Cursor, Windsurf or Zed. kino doctor lists which features are ready.
Integrations
206 tools with KINOCUT_TOOLS=all, 18 by default through eight operation families, plus 177 CLI commands and a Python Client. Optional extras add Whisper transcription, Demucs stems, Real-ESRGAN upscaling and object mattes; Hyperframes and Revideo handle motion graphics.
Data and privacy
Runs locally over stdio and opens no port. It reads any media path it is given and writes new files; relative traversal, symlinks and sensitive dotfiles are blocked, absolute paths are allowed. Vision features are off by default; remote mode sends frames to api.anthropic.com. The object-matte extra downloads about 1 GB of weights on first use.
Cost at scale
Free. Costs are local compute and model tokens: the default tool list is about 22 KB of JSON per session, the full catalogue about 178 KB. Remote vision QC, if enabled, bills your Anthropic key.
Limits
Renders real media only: no text-to-video generation, and caption speech uses eSpeak NG rather than a neural voice. video_shorts letterboxes widescreen sources for now. About 13 open issues; AI extras run to several GB.

What’s good

  • Never overwrites a source without explicit permission
  • 18 tools by default, 206 on demand
  • Quality gates for black frames, silence and loudness
  • No account, uploads or unrequested network calls

What’s not

  • Whisper captions need a large optional install
  • Absolute output paths can write anywhere
  • Open bug lets a failed-QC cut be exported
  • Rapid releases; pin the version

Best for

Developers, podcasters and small video teams who want an agent to cut and package clips on their own machine without risking the source footage.

Kinocut alternatives

Kinocut: common questions

Is Kinocut worth it?
A video server built around not wrecking your footage, with the deepest tool catalogue in its niche and a release cadence that asks you to pin versions. We rate it 3.9 out of 5.
How much does Kinocut cost?
Free and open source (Apache-2.0). No account, credits or hosted service.
What are the downsides of Kinocut?
Whisper captions need a large optional install; absolute output paths can write anywhere; open bug lets a failed-QC cut be exported; rapid releases; pin the version.
Who is Kinocut best for?
Developers, podcasters and small video teams who want an agent to cut and package clips on their own machine without risking the source footage.
How much does Kinocut cost for a team or in production?
Free. Costs are local compute and model tokens: the default tool list is about 22 KB of JSON per session, the full catalogue about 178 KB. Remote vision QC, if enabled, bills your Anthropic key.
Is Kinocut safe for company data?
Runs locally over stdio and opens no port. It reads any media path it is given and writes new files; relative traversal, symlinks and sensitive dotfiles are blocked, absolute paths are allowed. Vision features are off by default; remote mode sends frames to api.anthropic.com. The object-matte extra downloads about 1 GB of weights on first use.
What are the best Kinocut alternatives?
The closest alternatives we have reviewed are OpenScreen, OpenEdit and agentic-obs (OBS Studio MCP).

Answers checked Nov 11, 2026.