Cutbolt MCP server
Commit Oct 25, 2026A local video editing engine that allows AI agents to arrange shots, add captions, mix sound, and export footage using JSON commands or MCP.
Also known as skulitom/Cutbolt, cutbolt
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 0.4.0
- Last commit
- Oct 25, 2026
- License
- MIT
What is Cutbolt MCP?
Cutbolt is a local video editing engine that allows AI agents to arrange shots, add captions, mix sound, and export footage using JSON commands or MCP.
Summary generated from the project's README
About Cutbolt MCP server
Cutbolt is a local video editing engine designed for AI agents to process recorded footage. It allows users to transform raw files like interviews, tutorials, and screen recordings into finished products by arranging shots, adding captions, mixing audio, and exporting files. The engine runs entirely on your machine via MCP over stdio, a Rust library, or JSON commands, ensuring that source media remains private and no cloud accounts are required.
Developers can integrate this tool with Claude Code or other MCP clients to automate complex editing tasks. The system maintains a revision history for every saved edit, which enables an agent to inspect changes, retry failed requests, or undo specific decisions. To function, the environment must have FFmpeg and ffprobe available on the system path. While it supports H.264 and HEVC footage directly, it can also handle other sources during the export process. The engine is currently verified on Windows and Linux platforms.
What you can do with it
- 1Automated interview editingAn agent can shorten pauses and switch between synchronized camera angles in a talking head video.
- 2Tutorial productionYou can instruct an agent to combine screen recordings with camera footage and overlays for a presentation.
- 3Audio and caption managementThe engine allows an AI to mix audio tracks and insert timed captions into a video project.
- 4Iterative video refinementAn agent can review its own work and revert to previous versions if an edit is incorrect.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableskulitom/cutbolt is public and reachablegithub.com/skulitom/cutbolt ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-25github.com/skulitom/cutbolt ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubCutbolt
Cutbolt is a local video editing engine for AI agents. Turn recorded footage into a finished edit: arrange shots, tighten dialogue, add captions and graphics, mix sound, and export. Every saved edit has a revision, so an agent can inspect changes, retry a request or undo a decision.
Use it through JSON commands, a Rust library or MCP over stdio. The engine runs on your machine, preserves source media, and needs no account or hosted service.
- Who it's for: people who turn recorded footage (interviews, talking heads, tutorials, screen recordings and presentations) into finished videos and want Claude Code or another MCP agent to do the editing on their own machine.
- Use it when an agent should cut, caption, mix and export real footage with exact timing, and you want to inspect, retry or undo each change it makes.
- Don't use it when you want to edit by hand in a visual timeline (Cutbolt is driven by an agent, JSON commands or a Rust library), need a hosted or cloud service, want new video generated from a text prompt (it edits footage you already have), or need formats beyond its documented limits (see current scope).
- Compared with Kinocut (formerly mcp-video), claudeclip, ffmpeg-skill, Video Jungle MCP, Vlexor and HyperFrames in Local video editing MCP servers for AI agents, and hosted alternatives: 7 compared (re-checked 25 October 2026): Cutbolt measured over MCP, including the 0.4.0 release as downloaded (frame-exact cuts, safe retries, undo, byte-identical repeat exports on one FFmpeg build), ffmpeg-skill's cut run on the same clip, the others read from their code and docs, and when to choose each instead.
What you can make
- Interviews and talking heads: cut by time or transcript, shorten pauses, switch between synchronized cameras, and add timed captions.
- Tutorials and recorded presentations: combine screen recordings, camera footage, titles and overlays; preview a frame or an exact section before exporting.
- Layered edits: arrange video and audio tracks, reuse nested sequences, and save transforms, masks, color adjustments and transitions with the project.
- Finished deliveries: mix and measure audio, export a lossless master or H.264 video, review the result, and hand off a supported timeline through OTIO.
- Repeatable workflows: keep revisions, request receipts and undo history; resume local render jobs and revise individual parts of a production.
Eligible H.264 and HEVC footage can be edited directly. Other supported sources prepare automatically during queued export. Titles, graphics and captions can remain editable scene recipes on the timeline. Optional speech recognition and media workers run locally with separately installed dependencies.
Get started
Cutbolt is in development. Windows and Linux are the current verification platforms. Every install needs FFmpeg and ffprobe on PATH, or CUTBOLT_FFMPEG and CUTBOLT_FFPROBE set to them; Cutbolt bundles neither.
Prebuilt (early release): the latest release has the CLI for Windows x64, Linux x64 and macOS ARM64 (macOS is not a verification platform: background jobs, which MCP exports go through, need Windows or Linux, and Windows-only engine features stay Windows-only), and cutbolt-windows-x64.mcpb, an MCP bundle for Claude Desktop on Windows. Check downloads against SHA256SUMS. Cutbolt is in the official MCP Registry as io.github.skulitom/cutbolt.
From source, with Rust and a C build toolchain, in this repository's checkout:
cargo install --path . --locked
cutbolt --version
cutbolt doctor
Create a separate workspace for your footage and outputs, then configure your agent's MCP client to launch:
cutbolt --workspace /absolute/path/to/workspace mcp
Replace the path with your workspace directory; on Windows, use a path such as C:\Videos\MyProject. The default MCP catalog is compact, with detailed commands available on demand.
See installation and client setup for prerequisites, configuration and the optional Claude Code plugin. The plugin includes recipes for recuts, tutorials, multicam editing, captions and delivery. Other agents can start with llms.txt.
To use the Claude Code plugin instead of configuring the client by hand, add it from GitHub:
claude plugin marketplace add skulitom/Cutbolt
claude plugin install cutbolt@cutbolt-local
Then restart Claude Code in your footage workspace. The plugin starts cutbolt from PATH and installs nothing itself. Its recipes follow main, which can be ahead of the prebuilt release, so install the CLI from source for the plugin: as above, or with cargo install --git https://github.com/skulitom/Cutbolt --locked.
Once connected, try a request such as:
Inspect the footage in this workspace, trim the first clip to its opening ten seconds, and show me a preview. Save the edit so I can revise it later.
For direct CLI use, the usage guide describes JSON requests and saved projects. The agent interface covers revisions, retries, jobs and workspace boundaries.
Current scope
Version 0.4.0 is an actively developed engine. Supported formats and operations have explicit limits; the timeline uses 8-bit SDR processing by default. See direct footage, clip effects, generated clips and export formats for the current contracts.
Verification checks decoded pictures, audio samples, timing and saved-edit behavior against independent expectations. Published evidence records the last full acceptance run; progress history records subsequent changes, targeted checks and open work. Agent workflow measurements and footage benchmark results describe the tested tasks and their limits.
Development and license
See the development workflow, architecture and 1.0 roadmap for implementation and planning details.
Original project code is licensed under MIT. External dependencies retain their own licenses; tested versions and requirements are in the dependency ledger. Repository material and attribution rules are documented in research and publication boundaries.
Sibling tools
Cutbolt is one of five free, MIT-licensed tools for agents by the same author. Each runs locally as an MCP server:
- Anode: Windows desktop automation in a hidden background session, so an agent can click, type and take screenshots without taking over your screen.
- Pixelbolt: pixel art and sprite animation from editable JSON recipes, exported as sprite sheets, atlases, animated PNGs and GIFs.
- Inkbolt: vector and raster image editing for diagrams and illustrations, with SVG, PNG, JPEG and TIFF import and SVG and PDF export.
- Nestbolt: 2D nesting of parts on sheets for laser, CNC, plasma and waterjet cutting, also in your browser.
Cutbolt MCP questions
What kind of video editing can Cutbolt perform?
It can arrange shots, tighten dialogue, add captions and graphics, mix sound, and export finished videos like interviews or tutorials.
Does Cutbolt require a cloud subscription or account?
No, the engine runs locally on your machine and requires no account or hosted service.
What are the system requirements for running Cutbolt?
You must have FFmpeg and ffprobe on your PATH, or manually set the CUTBOLT_FFMPEG and CUTBOLT_FFPROBE environment variables.
Can I undo changes made by an AI agent?
Yes, every saved edit includes a revision, allowing an agent to undo decisions or retry requests.