cuttingfloor changelog

version

0.43.1 → 0.45.7

version

0.43.0 → 0.43.1

version

0.42.0 → 0.43.0

version

0.41.0 → 0.42.0

version

0.40.1 → 0.41.0

version

0.39.0 → 0.40.1

version

0.38.1 → 0.39.0

version

0.37.1 → 0.38.1

version

0.37.0 → 0.37.1

version

0.36.1 → 0.37.0

version

0.36.0 → 0.36.1

version

0.34.12 → 0.36.0

version

0.34.11 → 0.34.12

version

0.34.9 → 0.34.11

version

0.34.8 → 0.34.9

version

0.34.4 → 0.34.8

https://cuttingfloor.ai/llms.txt

− > Turn Claude, Cursor or ChatGPT into a video editing agent. One MCP server, 62 tools: cut by transcript, caption, mix and export real video. Open source.− cuttingfloor is a headless video editor and video editing MCP server. Connect an AI agent (Kazi, Claude, Claude Code, Cursor, ChatGPT in developer mode, VS Code) and it becomes a video editing agent: it imports footage, transcribes it, cuts on words, adds captions, mixes audio to a loudness target, makes graphics and voiceovers, checks its work and exports an MP4.− ## Connect− - MCP server: https://api.cuttingfloor.ai/mcp (Streamable HTTP, OAuth sign-in)− - Claude Code: `claude mcp add --transport http cuttingfloor https://api.cuttingfloor.ai/mcp`− - Cursor: `{ "mcpServers": { "cuttingfloor": { "url": "https://api.cuttingfloor.ai/mcp" } } }` in ~/.cursor/mcp.json− - Claude: Settings → Connectors → Add custom connector− - ChatGPT: Settings → Apps → Create (developer mode), auth OAuth− ## Links− - [Home](https://cuttingfloor.ai): overview, pricing and sign-up− - [Source and docs](https://github.com/mvp2o-ai/cuttingfloor): open source, runs in one Docker image− - [Skills](https://cuttingfloor.ai/skills): step-by-step instructions an agent follows with cuttingfloor's tools− - [Demos](https://cuttingfloor.ai/demos): real clips edited by an agent with cuttingfloor connected, before and after− ## Demos− - [Cut a dance video to the beat, with no copyrighted music](https://cuttingfloor.ai/demos/edit-a-dance-short-to-the-beat): A dancer gives Kazi four takes of a Motion Party-style routine. Kazi makes an original beat, cuttingfloor cuts a new shot on every bar, slows into the big move and puts each move’s name on screen: a 23-second Short.− - [Edit a two-person interview into a Short](https://cuttingfloor.ai/demos/edit-a-two-person-interview): A podcast host asks Claude to cut six camera angles of an interview together. cuttingfloor cuts to whoever is talking, removes the dead air and the ums, adds name cards and captions, and makes vertical and widescreen versions.− - [Turn a beauty product review into a 15-second Short](https://cuttingfloor.ai/demos/edit-a-product-review-short): A beauty creator asks ChatGPT to make a lip gloss review into a Short. cuttingfloor opens on the best line, speeds through the application, pops up callouts as she says them and cuts her out onto a pink end card with the price.− - [Edit a real estate listing video with ChatGPT](https://cuttingfloor.ai/demos/edit-a-real-estate-listing-video): A realtor asks ChatGPT to cut a wrong address and the dead air from a listing clip, then add captions and a price card. cuttingfloor turns 20 seconds into a 14-second post.− - [Remove the pauses and restarts from a talking video](https://cuttingfloor.ai/demos/remove-pauses-and-restarts-from-a-talking-video): A creator asks Claude to cut the long pauses, the restart and the repeated line out of a 40-second tips video. cuttingfloor keeps every point, adds big word-by-word captions and a hook title: a 21-second Short.− - [Let AI pick your best take, then make the thumbnail](https://cuttingfloor.ai/demos/pick-your-best-take-with-ai): A tech creator gives ChatGPT four takes of his channel intro. cuttingfloor has Jev, its own vision model, watch every quarter second of each take, keeps the best one, then picks the frame where he looks best for a thumbnail.− - [Add a talking head to a slide explainer, then make it a Short](https://cuttingfloor.ai/demos/add-a-talking-head-to-your-explainer-video): A creator asks Claude to put himself into Tip 1 of his 24-minute slide explainer. cuttingfloor finds the tip by its words, cuts him out of his background into the corner the slides leave empty, then makes a vertical Short: slides on top, him underneath, captions on the seam.− ## Skills− A connected agent finds these itself: skills.search {query} in the user's words, then skills.get {slug} for the whole skill.− - [Cut pauses, ums and restarts](https://cuttingfloor.ai/skills/cut-pauses-ums-and-restarts): Clean up a talking video with cuttingfloor. Cut the long pauses, the ums, the lines the speaker started over, and lines said twice, then smooth every cut so nothing jumps. Use when someone asks to tighten a talking-head, vlog, interview, tutorial or listing video, remove dead air, remove filler words, or fix restarts and retakes.− - Trial: $0 3 days, 60 render min. One job at a time, 1080p, 5 GB storage. Everything, no card, for 3 days. About six hours of an agent editing.− - Creator: $49 / month, 300 render min a month. One job at a time, 1080p, 50 GB storage. About 30 hours of editing a month. One job at a time, up to 1080p.− - Pro: $99 / month, 900 render min a month. 2 jobs at once, 4K, 200 GB storage. Agents editing most days. Two jobs at once, up to 4K.− - Enterprise: Talk to us, From 5,000 min a seat. Priority, more jobs at once, 4K. Agents editing all day, for a team. Minutes, seats and dedicated servers to suit.− A render minute is one minute of a job running (export, import, transcript, graphics, voiceover, cutout). Editing the timeline is free. At a plan's minutes new jobs wait for the renewal; nothing is billed past the price.− ## Compared with alternatives− - cuttingfloor: An MCP server of 62 editing tools. The agent imports footage, reads the transcript, cuts, captions, mixes and checks every step with a preview. Pick it when the agent decides the edit, on footage it has not seen before.− - Remotion: Videos written as React components and rendered from code. Pick it when you want the video to be a codebase you keep and maintain.− - Shotstack: A cloud render API: send a JSON timeline, get a file back. Pick it when you already know the edit and need it rendered at volume.− - Creatomate: Templates designed in its editor, filled in through an API. Pick it when one branded template needs many variations.− - JSON2Video: A JSON-to-video API with built-in text-to-speech. Pick it when simple videos are assembled from data.− ## FAQ− ### Can Claude edit videos?− Yes, with cuttingfloor connected. Claude (web, desktop and Claude Code) calls its tools to import your footage, transcribe it, cut on words, add captions, mix the sound and export an MP4. In Claude, add it under Settings → Connectors; in Claude Code, run claude mcp add.− ### Can ChatGPT edit videos?− Yes, in developer mode, which lets ChatGPT use remote MCP servers. Add the cuttingfloor address under Settings → Apps → Create, sign in, and ask for the cut you want.+ > Editing tools for the agent you already use. Connect Claude, ChatGPT, Codex, Cursor or Kazi to one MCP server, 72 tools: it cuts by transcript, captions, mixes, makes graphics and exports real video, checking its work as it goes. Open source: run it yourself free, or hosted.+ cuttingfloor is video editing tools for your agent, not an editor with an AI inside. It is an open-source MCP server: your own agent (Kazi, Claude, ChatGPT, Codex, Cursor or any MCP client) calls its granular tools to import footage, transcribe it, cut on words, caption, mix to a loudness target, make graphics and voiceovers, check its work with previews and export an MP4, FCPXML, OTIO or EDL. The judgment is the agent's; cuttingfloor makes the edits and reports facts (transcripts, speech edges,+ ## Run it+ - Hosted: `https://api.cuttingfloor.ai/mcp` (Streamable HTTP, OAuth sign-in). Free 3-day trial, no card.+ - Self-hosted: free and open source (AGPL-3.0), one Docker image (`ghcr.io/mvp2o-ai/cuttingfloor`): `git clone https://github.com/mvp2o-ai/cuttingfloor && cd cuttingfloor && docker compose up -d`, then connect to `http://127.0.0.1:8787/mcp`. [Self-host](https://cuttingfloor.ai/self-host) ([Markdown](https://cuttingfloor.ai/self-host.md))+ - Trial: $0 for 3 days, 60 render min. One job at a time, 1080p, 5 GB storage.+ - Creator: $49 / month, 300 render min a month. One job at a time, 1080p, 50 GB storage.+ - Pro: $99 / month, 900 render min a month. 2 jobs at once, 4K, 200 GB storage.+ - Enterprise: Talk to us, Minutes and limits agreed on a call. Start on any plan now; move up when we have talked.+ - Self-hosted: free (AGPL-3.0), no minute limits.+ A render minute is one minute of a job running on the hosted servers (export, import, transcript, motion graphics, voiceover, cutout). Planning and editing the timeline are free. When a plan’s minutes are used, new jobs wait for the renewal or a bigger plan; nothing is ever billed past the plan price. Details: [Pricing](https://cuttingfloor.ai/pricing) ([Markdown](https://cuttingfloor.ai/pricing.md)), [pricing.json](https://cuttingfloor.ai/pricing.json)+ ## What needs your own key+ | Feature | Tools | Hosted | Self-hosted |+ | --- | --- | --- | --- |+ | Lip-sync (redraw a speaker’s mouth to new words) | `create.lipsync` | Your own sync. labs key, saved on your account page or sent as a header. sync. labs bills you per second. | Your own sync. labs key as SYNC_API_KEY. |+ | Visual judgment: questions about frames, looking across a video, finding moments in plain words | `judge.ask`, `media.analyze (look)`, `media.seek` | Included. No key. | A TypeSafe key as TYPESAFE_API_KEY (typesafe.ai). |+ Everything else needs no key, hosted or self-hosted; it is built into the image and runs on the server:+ - Transcripts with word times (whisper.cpp)+ - Voiceover with timed words (Kokoro)+ - Background removal for people or any subject (MODNet, BiRefNet)+ - Voice, drums, bass and music split apart (Demucs)+ - Stabilizing shaky footage (vid.stab)+ - Motion graphics and stills from HTML, with transparency+ - Beats, faces, object tracking, silences and scene changes+ - Every edit, effect, caption, mix and export (FFmpeg and MLT)+ ## Connect+ - **Kazi**: In Kazi (app.apifunnel.ai), open Servers and add your own server. Paste the address, Connect, then sign in. `https://api.cuttingfloor.ai/mcp`+ - **Claude**: Customize → Connectors → + → Add custom connector. Paste the address, then sign in. `https://api.cuttingfloor.ai/mcp`+ - **Claude Cowork**: Add cuttingfloor as a connector in Claude, as above. Cowork uses the same connectors as your Claude account. Start a Cowork task and ask for the edit. It calls cuttingfloor’s tools as it works.+ - **Claude Code**: Run this, then /mcp to sign in. `claude mcp add --transport http cuttingfloor https://api.cuttingfloor.ai/mcp`+ - **ChatGPT**: chatgpt.com/plugins → + → Add custom MCP server. Paste the address, authentication OAuth. `https://api.cuttingfloor.ai/mcp`+ - **ChatGPT Work**: Add cuttingfloor as a plugin in ChatGPT, as above. Start a Work task and type @cuttingfloor with the edit you want.+ - **Codex**: Run this, then codex mcp login cuttingfloor to sign in. `codex mcp add cuttingfloor --url https://api.cuttingfloor.ai/mcp`+ - **Cursor**: Add to ~/.cursor/mcp.json. Cursor opens the sign-in when you first use it. `{ "mcpServers": { "cuttingfloor": { "url": "https://api.cuttingfloor.ai/mcp" } } }`+ - **VS Code / GitHub Copilot**: Open the Command Palette and run MCP: Add Server. Choose HTTP, paste the address, name it cuttingfloor and pick Global. Start it and sign in when VS Code asks.+ - **Antigravity CLI (Gemini)**: Add cuttingfloor to ~/.gemini/config/mcp_config.json (or .agents/mcp_config.json in a project). Run /mcp to check it’s there, and sign in when asked. In the Antigravity editor: … → MCP Servers → Manage MCP Servers → View raw config, and add the same lines.+ - **Zed**: Open Settings → AI → MCP Servers → Add Server → Add Remote Server, and paste the address. Or add the lines below to your settings.json. Zed opens the sign-in for you.+ - **Goose**: Click Add to Goose below, or in the desktop app open Extensions → Add custom extension. In the CLI, run goose configure → Add Extension → Remote Extension (Streamable HTTP), and paste the address. Goose opens the sign-in for you.+ - **Any MCP client**: Add a remote server at https://api.cuttingfloor.ai/mcp. It speaks Streamable HTTP. Your app opens the cuttingfloor sign-in page (OAuth). Sign in, and it’s connected. For a script with no browser, make a key on your account page (it starts with cf_) and send it as a header instead.+ Self-hosted, use the local address instead of the hosted one. Server card: https://api.cuttingfloor.ai/mcp/server-card · Catalog: https://cuttingfloor.ai/.well-known/ai-catalog.json

Showing the first 40 of each.

https://raw.githubusercontent.com/mvp2o-ai/cuttingfloor/HEAD/AGENTS.md

− - Promote: `scripts/promote-develop.sh "<summary>"` opens the develop → main pull request, waits for its checks and merges it. It versions pending changesets on `develop` first; on `main` the release workflow tags the version and publishes the image. `main` takes nothing but that pull request. Only `develop` may open a pull request into `main` (the Flow check).+ - Tools are agnostic: never name or shape a tool after the thing it acts on or the job it serves. "Remove the logo", "blur the dog" and "brighten the mug" are one chain, as in Photoshop: select (a lasso or shape mask, or an item by its bounds), then act on the selection (show, hide, grade, blur, fill). Bounds and other inputs come from tools that already exist (`track`, `faces`, `judge.ask` boxes). A gap is the one missing verb or fact in such a chain, named for what it does and what it takes (a+ - Promote: `scripts/promote-develop.sh "<summary>"` opens the develop → main pull request, waits for its required checks (not the image build, which gates the server deploy on `main` and skips site-only changes) and merges it. It versions pending changesets on `develop` first; on `main` the release workflow tags the version and publishes the image. `main` takes nothing but that pull request. Only `develop` may open a pull request into `main` (the Flow check).

version

0.34.0 → 0.34.4

version

0.32.1 → 0.34.0

• + editor
• + library

• + create.upscaled

version

0.31.0 → 0.32.1

description

Video editing for AI agents: cuts, captions, audio, motion graphics, previews, and exports. → Hosted video editing for AI agents: cut, caption, mix and export real video. Sign in with OAuth.

version

0.19.2 → 0.31.0

https://cuttingfloor.ai/llms.txt

− - [Cut a dance video to the beat, with no copyrighted music](https://cuttingfloor.ai/demos/edit-a-dance-short-to-the-beat): A dancer gives Kazi four takes of a Motion Party-style routine. cuttingfloor makes an original beat, cuts a new shot on every bar, slows into the big move and puts each move’s name on screen: a 23-second Short.− - Trial: $0 3 days, 60 render min. One job at a time, 1080p. Everything, no card, for 3 days. About six hours of an agent editing.− - Creator: $49 / month, 300 render min a month. One job at a time, 1080p. About 30 hours of editing a month. One job at a time, up to 1080p.− - Pro: $99 / month, 900 render min a month. 2 jobs at once, 4K. Agents editing most days. Two jobs at once, up to 4K.− cuttingfloor is one: 62 tools over MCP, from clip.add and caption.add to audio.set_master and export.video. Every edit returns a new timeline and a preview, so the agent sees what it did and can step back. When an edit needs a loop, code.run runs a script that calls the same tools.+ - [Cut a dance video to the beat, with no copyrighted music](https://cuttingfloor.ai/demos/edit-a-dance-short-to-the-beat): A dancer gives Kazi four takes of a Motion Party-style routine. Kazi makes an original beat, cuttingfloor cuts a new shot on every bar, slows into the big move and puts each move’s name on screen: a 23-second Short.+ - Trial: $0 3 days, 60 render min. One job at a time, 1080p, 5 GB storage. Everything, no card, for 3 days. About six hours of an agent editing.+ - Creator: $49 / month, 300 render min a month. One job at a time, 1080p, 50 GB storage. About 30 hours of editing a month. One job at a time, up to 1080p.+ - Pro: $99 / month, 900 render min a month. 2 jobs at once, 4K, 200 GB storage. Agents editing most days. Two jobs at once, up to 4K.+ cuttingfloor is one: 62 tools over MCP, from clip.add and caption.add to audio.set_master and export.video. Every edit returns a new timeline and a preview, so the agent sees what it did and can step back.

https://raw.githubusercontent.com/mvp2o-ai/cuttingfloor/HEAD/AGENTS.md

− - Core tenet: tools give the agent actions (edits) and facts it cannot get itself (a transcript, speech spans, frames, loudness); judgment and composition are the agent's. A composition that isn't obvious goes in a skill (`skills/`), a soft harness of steps using the tools. It becomes a script or a tool only on evidence: it is common, agents measurably struggle with the skill, and the tool brings something the agent cannot reproduce. A tool never merges sources or decides behind the agent's back− - Every tool declares its output schema, and results always match it (including the job a long-running tool returns). Agents chain calls in `code.run` before any result exists, so the output signature is the contract; the generated SDK carries it.− - No provider key ever appears in a tool argument, result, error, log or stored file.+ - Core tenet: tools give the agent actions (edits) and facts it cannot get itself (a transcript, speech spans, frames, loudness); judgment and composition are the agent's. A composition that isn't obvious goes in a skill (`skills/`), a soft harness of steps using the tools. It becomes a script or a tool only on evidence: it is common, agents measurably struggle with the skill, and the tool brings something the agent cannot reproduce. A tool never merges sources or decides behind the agent's back+ - Every tool declares its output schema, and results always match it (including the job a long-running tool returns). Clients chain calls (in their own code execution) before any result exists, so the output signature is the contract. cuttingfloor runs no client code: running MCP servers from code is the client's job.+ - No provider key ever appears in a tool argument, result, error, log or stored file. A key the user saves on their account page (their sync. labs key) is kept only as ciphertext (AES-256-GCM, bound to the user, with a server-only secret) in their WorkOS metadata, never on the server's disk, and only its last four characters are shown.