Launching today
Cuthol
Open-source video editor built for AI agents
2 followers
Open-source video editor built for AI agents
2 followers
Tell your AI agent what you want and it edits the video: cuts pauses and retakes, captions every word, ducks music under the voice, adds code-based motion graphics, checks its own work and renders. Local, MIT-licensed, driven through a CLI and an MCP server.
Free
Launch Team / Built With
Render Sandboxes: a computer for every agentIsolated compute, persistent state and a real filesystem
Promoted
Hi Product Hunt,
We're HypeAI. We know that people—even professional editors—are now using AI models to perform editing tasks. For this reason, we are open-sourcing our tool, which is designed to position the AI as the user of the editing software rather than the editor itself: Cuthol, the video editor built for AI agents.
Why we built it Agents can already write code, but editing software is designed for a person with a mouse and a visible timeline, and an agent has neither. The usual workaround is to have the agent generate the whole video as code. That works for a short clip, but every late change means regenerating the composition, and on a long voice track that gets slow, expensive and fragile. We wanted an agent to edit the way an editor does: on a persistent timeline, change by change, with the ability to check its own work.
What Cuthol is A local, open-source (MIT) non-linear editor that your agent drives through a CLI, an MCP server (14 tools) or skills. It works with Claude Code, Codex, Cursor, Gemini CLI and any MCP client. The agent works through typed commands, reads the results as text, looks at real preview frames when it needs to, and every change can be undone.
What it can do • Timeline: a magnetic primary storyline plus typed lanes for titles, captions, b-roll, motion graphics, voiceover, sound effects and music. Every clip and cue has a stable id, so the agent never edits the wrong thing after a change. • Speech: transcription with word timings, pause and filler cleanup that cuts only in real silence, retakes flagged for review instead of cut blindly, word-highlighted captions and voiceover. • Mix: loudness normalization (e.g. −14 LUFS), voice polish, ducking under speech and cut alignment to a beat grid. • Motion graphics: HyperFrames compositions composited with true alpha, fed with the edit's word times, beats and cuts. HTML, GSAP and Three.js 3D render frame-exact. • Seeing and measuring the edit: frame previews and contact sheets from the same pipeline as the final render, plus per-frame geometry of every graphic element, so layout problems are checked in text. • Safety: an atomic journal with undo previews, lane locking, five verification gates before every render and detection of files changed outside Cuthol. • Extensions: agents add their own tools without rebuilding Cuthol: composition helpers, asset packs, processors, effects, transitions, planners and speech/music providers. • Render: no time limit, overlay caching, parallel overlay capture and GPU encoding with CPU fallback.
What it did on a real production We measured one 6:35 video on our channel from the agent's own session logs: • Extensions compound. On the same 3D desk set, cost per minute of finished video went from ~$29 for the first scene built from scratch to ~$6.38, even counting every 3D model at full price (agent spend alone: ~$2.36). That is 4.6× to 12× cheaper, and it improves with every video. • Edit, don't regenerate. Cutting one sentence from the finished edit was a one-line change and a ~2-minute rebuild. 525 sound cues and 49 camera moves re-timed themselves, and voice sync held at 0.00 s drift across 56 of 56 checkpoints. • The agent checks its own work. It inspected 134 frames, fixed 12 issues before a preview was rendered and had 2 errors stopped before a render started. (Costs are API-equivalent at Claude Opus 5.5 list prices; the method is in our video description.)
What feedback would help most
Which agent or MCP client do you use, and where did setup slow you down?
Which editing task would you hand to an agent first?
Which extensions or providers should come next (local speech models, stock music sources, new effects)?
How does rendering perform on your hardware?
Cuthol runs entirely on your machine and your footage never leaves it. Try it, break it, and tell us what you'd build with it. We'll be here all day answering questions.
Thank you for checking it out!