Built by AI agents. Steered by a person.
OpenFrame is an experiment in autonomous software development: can AI coding agents build a professional, fully open video editor? Agents wrote almost all of it, most of it Claude Opus 5.5 working in Claude Code, with Codex alongside. A human owner set the direction, tried the builds and decided what shipped.
Who did what
The agents
- Researched CapCut, Premiere, Resolve and VEGAS before building each part.
- Wrote the code, the tests and the documentation.
- Ran the checks and drove the running app in a headless browser to try their changes.
- Recorded every non-obvious decision with its reasons: 116 so far.
- Reviewed each other's work: Claude Code reviewed the Codex pass.
The owner
- Chose what to build: desktop first, everything on the user's hardware, GPL-3.0-only, the name.
- Tried the builds and gave feedback on what felt wrong.
- Decided what to keep, when to open the repository, and when to publish.
- Holds the rule that generated work is not correct until it is checked, and that the record says how.
How agents work on it.
An agent forgets everything between sessions. The project is set up so the next one does not have to.
-
Documentation is the memory
Architecture decision records, research notes, a history of every round and a handoff for whoever comes next. Agents read them before they change anything.
-
Parallel worktrees
Several agents work at once, each in its own git worktree, and their work is merged when it is finished and checked.
-
Gates before every commit
Type checks, unit tests, and Rust formatting, lints and tests. CI runs the same gates when the owner starts it.
-
Checked in the running product
Tests are not enough: agents drive the real app, headless, through the same automation surface the MCP tools use, and look at the frames.
The build so far.
The first entry in the project's history is October 13, 2026. These are the rounds, as the history records them.
-
The project starts
The owner asks for a modern CapCut with VEGAS and Premiere depth. Five research passes study CapCut, professional editors, the browser media stack, Tauri and AI before any engine code. The core is written test-first, and every edit becomes a named command with a schema before most of the interface exists, so agents could use it from day one. The owner sets the course: desktop first, all compute on the user's hardware.
-
Agents in parallel
Agents build in their own git worktrees and merge one by one: the desktop platform, a frame-exact export, and more. The owner chooses the license: GPL-3.0-only.
-
Depth, reviewed
Codex takes a pass and Claude Code reviews it. The timeline's interactions, crop and masks on the canvas, compound clips, a CapCut feature-parity pass and background removal land.
-
OpenFrame, in public
The app is renamed OpenFrame and the repository opens. Claude makes a whole promo through the MCP tools, and what it struggled with reshapes them: smaller answers, batches, contact sheets.
-
Tools for professional video
Brand kits, ad elements, kinetic type, deliverables and sound design for agents. The launch film is edited through MCP, its limits are written down, and the next round fixes them.
The films were made by agents too.
The launch film and every clip on this site came from agents working with OpenFrame itself.
The launch film was cut, scored and narrated by an AI agent inside OpenFrame, through its MCP tools. The only part made outside the editor is the footage of the app itself.
The footage
An agent built demo projects through the MCP tools from public-domain NASA footage and CC0 clips from Wikimedia Commons, then recorded takes of the web build in a headless browser, with a drawn cursor, undoing each take's edits so every take started from the same project. The film uses 18 takes recorded this way from the current build.
The score
The agent planned a 120 BPM score in F minor against the edit before anything rendered: drops at 0:08 and 0:40, a breakdown before the second, and the final hit at 1:00, where the end card lands. Every asked time fell on a bar line. OpenFrame's music generator rendered the 68 seconds in about five seconds, mastered to -14 LUFS, synthesizing every instrument in code with no samples.
The voice
The narration is 14 lines read by OpenFrame's text to speech on the computer, each placed on its beat, with a pronunciation map. Then the agent had the app listen back: Whisper heard every word, and the review named three problems with exact fixes, a misheard word, two slow lines and music too loud under the voice. Three lines were read again, and the music was ducked to sit 12 LU under the voice.
The edit
The master took about 30 MCP calls, most of them batches, one per section: kinetic type with styled accent words, floating windows with shadows, tilt and motion blur, push-ins on the feature in use, soft glows drawn as radial gradients, and whooshes on the reveals. This time no sprite or music was made outside the editor. The agent checked each section on contact sheets and ran the project check until it passed with no warnings, after it caught text under the size floor, contrast as low as 1.27:1 and headlines past the title-safe edge. The export is at -14 LUFS, with the narration's captions beside it, the ones this site's player offers.
The vertical cut
A 9:16 version for TikTok, Reels and Shorts, with the captions burned in, came from a copy of the project. The versions tool could not restack this film's side-by-side designs for a tall frame, so the agent laid it out again by hand in about 25 steps and checked it against the parts of the frame those apps cover.
What the films taught the editor
Each agent listed every limit it met. The work since the last film fixed the ones that misled it:
- Keyframes on properties that do not animate were accepted silently; now they are refused, and answers show the keys (ADR-066).
- Other shapes of a hand-animated film pushed its windows off frame; versions now lay keyframed motion out again (ADR-067).
- The project check guessed at contrast it never saw; it now measures the picture and the sound it judges (ADR-068).
- Music analysis found one drop and missed a breakdown; it now finds every drop and measures voice balance (ADR-075).
- The music came from outside the editor; now an agent can compose the score in it (ADR-085).
This film's list is the next round's work: laying side-by-side designs out again for 9:16, layouts that can fit a picture instead of cropping it, a title edit that dropped a styled word, and new clips that replaced others without saying so. One fix landed the same day: applying a layout no longer ripples the main track. The full notes
This website
An agent built this site too: the clips are cut from an agent's recordings of the running app, the screenshots come from a script that drives the running app and builds its demo projects through MCP, and the tool reference is generated from the app's own documentation every time the site is built.
What it is not, yet.
An honest list, from the project's own roadmap and release notes.
- Not packaged. There are no installers or signed builds: you build it from source.
- Not tested everywhere. The desktop app is developed on Windows; macOS and Linux builds are untested. The web build has been checked in Edge.
- Not every desktop permission is checked. Some features verified in the web build, such as camera and screen recording, still wait for a desktop check.
- Not finished. The roadmap lists what is next, from proxies on the web to multicam.