An AI video editor that proves its cuts. Your recordings in, a finished, mastered film out: cut by transcript, with b-roll, cards, music and captions, and every step an agent can call. proofcut renders on your own machine, then transcribes the render and checks that it says what the edit says.
https://github.com/user-attachments/assets/4153d180-3d7c-4c70-af5f-54d63d0a8bd5
An agent cutting a demo video, unattended, from a goal and no steps.
Silent, because the recorder took frames only; the voice is in the film the agent cut. The workspace (2:26) and Claude Code with the proofcut plugin (3:09).
Runs on Linux, where it is developed. macOS and Windows pass the test suite on CI; how far each has gone by hand is § Help wanted.
Quickstart
With uv installed:
uvx proofcut doctor # only looks: what is missing, and the fix for each
uvx proofcut setup --plan # what an install would fetch, and its size; touches nothing
uvx proofcut setup # installs it for you alone, no sudo; --uninstall reverses it
Then make the two-minute demo, or give an agent the tools with the Claude Code plugin. § Try it walks all of it.
Why it exists
Most of the work in a narrated video, whether an essay, a tutorial, a screencast or a talk, is bookkeeping. Find the retakes and cut them clean. Put the right footage under each line. Level the music under the voice, caption it, end on a card, render it, master it. An agent can do that bookkeeping now, given an editor it can drive.
What an agent cannot do on its own is know that the file it rendered is the film it meant. ffmpeg, melt and auto-editor all exit 0 on some failures, so a render can drop a line, keep a retake, add a frame of black or carry no captions and still report success. Every check that reads the project rather than the file agrees with it. The usual way to find out is to watch the whole thing.
proofcut is an editor built around that gap. You, or an agent, edit a timeline addressed by the words in it. proofcut renders it on your machine, then transcribes the render, counts its frames, and says where the file and the edit disagree.
Here it is on the demo project. The retake is cut from the edit, and
verify is pointed at a render made before the cut:
proofcut -C proj cut vo 11:23 --pad 0.1
proofcut -C proj verify stale.mp4
"expected_words": 34,
"heard_words": 47,
"similarity": 0.84,
"diff": ["--- timeline", "+++ render", "@@ -10,4 +10,17 @@",
" video", " editor",
"+every", "+cut", "+you", "+make", "+names", "+a", "+um", "+no",
"+let", "+me", "+try", "+that", "+again",
" every", " cut"]
The file still says the thirteen words the edit took out, and verify names
them. A render of the edit as it stands hears all 34
(DEMO.md
step 7).
Run, not staged
TRIAL.md scores three unattended runs, each handed a goal and no steps, and each passed every one of its checks:
- The demo cut, the one above.
- Real footage. 96 seconds of narration with its fluffed takes left in, plus four clips of film footage. The agent cut it to 45 seconds, chose footage by what each line was about, burned captions, and checked its own render: all 123 expected words heard back.
- A whole film. The demo material plus a score, briefed as a finished film ready to upload. In 192 seconds and $2.27 the agent cut it, laid the music under the voice, ended on a card, mastered it to −16 LUFS, and checked it. The score, the level and the end card were each measured in the delivered file, not taken from the project.
How it compares
COMPETITORS.md reads every entry from its source, most recently on 2026-09-28. In short:
- Transcript editors (the Descript model, and its open-source re-implementations) cut by deleting words, for a person's hands. None of the open ones has an agent surface, and none checks its render.
- Agent editors over MCP let an agent drive the edit. The nearest render checks among them compare the duration within a tolerance, or transcribe short windows around suspect seams. None diffs the whole render's words against the cut.
- NLEs (Kdenlive, Resolve, Premiere) are where a person finishes by hand,
and DaVinci Resolve 21.1 now ships an MCP server of its own. proofcut
exports to them rather than replacing them, and, with the trim brought back
by
import-edit,verifychecks what they deliver. - auto-editor removes silences headlessly. proofcut uses it as a render backend.
What proofcut lacks that others have: a check for frozen or silent spans in a render.
What it can do
Every stage from the first import to the delivered file is a proofcut command. No NLE finishes the film, and nothing else touches the render. Each row links the manual section that walks it.
Two video essays of five to six minutes, first finished in Kdenlive, have been rebuilt with these commands alone and measured against the delivered files. One came out the same length to the frame. The other matched all 63 of its voiceover ranges to the millisecond, with the voice aligned to the sample. Both master at the original's −16 LUFS. The measurements are in HISTORY.md.
Beyond those stages:
- Cuts stay addressable. A word's index never renumbers, so
cut vo 111:114names the same words however many cuts came before it, and a cue hung off a phrase stays addressed through every cut around it. - An agent that can look.
shot-sheetdraws the whole picture track as one labelled grid, andfootage-sheetbrowses a clip you haven't cut yet. Both return the image itself over MCP, not a path the agent can't open. - B-roll by description.
describewrites what is on screen in each ~10-second window of footage, so an agent can choose a clip by what a line is about. - Transcript self-checks. Retake seams, invented words, swallowed repeats
and suspect durations are reported when a transcript is attached, and
unspokenlets the render itself testify to words nobody said. - Reframing for another aspect. Per-shot crop windows, face-aware
proposals (
reframe-detect), a review sheet, and stacked splits for two speakers. Cards are redrawn at the new frame size, never stretched. - Screen recordings. Named instants (
events) that a crop, a sound or a speed change can hang off, a clip inset into the recording, and a retime for the slow parts. - Animated graphics. A web page, written by an agent or filled from a template (typing, a highlighter sweep, a letter build, chips), captured frame by frame and placed over a span of words. It animates in, holds for the span, and animates out, so a cut never breaks one. A page can show the project's own footage, exact to the frame, and saved graphics carry across projects.
- Stills and stickers. Photos and cut-outs go in once, upright, and play full frame or ride over the film: placed, sized, turned, framed as a photo card, popping or sliding in. Paste one into the agent's prompt and it is added.
- Built for agents. 132 MCP tools with typed inputs, structured returns and read/write annotations, so Claude Code, Codex or your own agent can drive it and a permission layer can tell a look from a change.
- No lock-in. The timeline is OpenTimelineIO, the manifest is JSON, and
the render is ffmpeg's and MLT's. Export a
.kdenliveor OTIO file and finish anywhere.
How people use it
Tell an agent what film you want. In Claude Code, install the plugin and describe the film: which recording is the voice, what the b-roll is, where the music goes, what it ends on. The agent imports, transcribes, cuts the retakes, hangs footage off the lines it belongs to, scores, masters and checks the render, using proofcut's tools and nothing else. It can look at the picture track as a labelled grid while it works, so it sees what it placed rather than a filename. Every MCP client runs the same server; the plugin is the one that needs no setup (§ In Claude Code).
Cut from a shell, by naming words. Every tool is also a proofcut
subcommand printing JSON, so a cut is a script you can read, re-run and
diff. --plan prints what a range says before anything changes, and undo
walks it back.
uv run proofcut -C myproject cut vo 111:114 --plan
Edit by hand in the workspace. Strike words in the transcript, drag-trim and razor on the timeline, review every crop in place, and export from Finish. It plays the source through the edit, so seeing a cut costs no render, and the truth strip warns while you edit if the film would ship wrong. The agent sits in a side rail of the same window.
uv run proofcut -C myproject open
Finish elsewhere, and still prove the file. Rough-cut here, export a
.kdenlive or OpenTimelineIO file, finish in Kdenlive, Resolve or Premiere,
and bring the trim back with import-edit. Then point verify at the
delivered file, and a retake left in is caught before anyone watches.
uv run proofcut -C myproject verify final.mp4
Cut a vertical teaser from the film. reel copies a span of the finished
film into a second project at another canvas, so the film itself is never
reshaped to take one render. It names every picture the teaser will not
have and pins the ones it keeps to the frames the film showed, so the
teaser shows what the film showed. Frame mode then reviews each crop on
the source's own frames.
uv run proofcut -C myproject reel ../teaser 1:32+44 --canvas 1080x1920
Split a pile of recordings into shorts. Seed every recording as one
timeline, cut the retakes once, then split it by first and last word. Each
short becomes its own project beside the pile, holding only the recordings
it uses, and any stretch no short took is named rather than lost.
uv run proofcut -C dump seed reel1 reel2 reel3
uv run proofcut -C dump split intro=reel1:0..reel1:46 demo=reel2:0..reel3:88
The first three are walked in § Try it; the reel, the split and the round-trip are in the manual.
What it holds to
- The file is the evidence. A check reads the render, never the project.
What
verify,framesandhold checkreport is measured off the delivered file, because a project can say captions are burned, or a line is in, about a file that has neither. - Your recordings, cut, not invented. proofcut makes no footage and writes no script. It takes what you recorded to a film, in an order you or your agent chose. It is not a text-to-video generator.
- Your machine, no meter. Commercial AI editors are apps around a metered cloud service. proofcut transcribes, edits and renders locally: no account, no per-minute billing, and no cloud service of its own. The only thing that talks to a model provider is the agent you choose to run, and your footage stays where it is.
- Nothing an agent does is out of your reach. The MCP server, the command
line and the workspace call the same operations, and a test holds every
tool to a matching command printing JSON. Changes take
--planto show what they would do first, anything addressed by word echoes the words it resolved to, andundo --steps Nwalks back an agent's whole turn. - Say what is not measured. The design record is public: the plans, and a dated history of what shipped and what the evidence said, failures included. Where only a CI runner has done something, the docs say a runner did it, not a person; where a check is a model's reading of a picture, it is reported and never gates.
Try it
Nothing below installs anything until you say so, and every step prints what it would do first. § What it puts on your machine is the whole footprint and how to reverse it.
Check your machine, before cloning anything. proofcut doctor probes every
tool proofcut uses (§ Requirements has the list) and prints
the fix for anything missing. It only looks, and writes nothing. With
uv installed:
uvx proofcut doctor
That first run downloads Python 3.13 if uv has none, plus proofcut's
dependencies. That is about 230 MB, all of it inside uv's own cache, which
uv cache clean empties.
Then read what an install would do. On Linux, Windows and either kind of
Mac, proofcut setup --plan prints every piece it would fetch, its size, and
which doctor row asked for it, then stops without touching anything:
uvx proofcut setup --plan
proofcut setup — installs into /home/you/.local/share/proofcut/deps
Will install
ffmpeg — n8.1.2 (BtbN autobuild-2026-08-31-13-27), about 126 MB
because: ffmpeg is not on PATH. ffprobe is not on PATH.
whisper — openai-whisper (uv tool, Python 3.12, cpu torch), about 1.9 GB
because: whisper is not on PATH.
auto-editor — 31.6.0, about 46 MB
because: auto-editor is not on PATH.
melt — Shotcut 26.8.1 (melt 7.41.0), about 155 MB
because: melt is not on PATH.
total: about 2.2 GB
That is a bare machine. Yours will be shorter, because setup installs nothing doctor passed, so a working ffmpeg or melt of your own is never touched. With an NVIDIA GPU the whisper row is CUDA torch and the total is about 5.8 GB.
Drop --plan to go ahead. It reprints the plan, asks once, and installs for
you alone, with no sudo or administrator rights; proofcut setup --uninstall
removes exactly what it added, and nothing you already had.
uvx proofcut setup
The demo and your own recordings run from a checkout:
git clone https://github.com/tydude001/proofcut && cd proofcut
uv sync
The two-minute demo
No footage needed. docs/DEMO.md makes a voiceover with a real retake, b-roll and a score, then walks a whole small film: cut the retake by naming its words, hang b-roll off a phrase, lay the score under the voice, render and master it, end on a card, and check the render against the timeline.
uv run python scripts/make_demo.py ~/proofcut-demo
In Claude Code
The plugin registers proofcut's MCP server, so every tool is available with no setup of your own:
/plugin marketplace add tydude001/proofcut
/plugin install proofcut@proofcut
The first start downloads about 175 MB of Python dependencies, and Claude
Code gives a server 30 seconds to connect. On a slow connection, start that
first session as MCP_TIMEOUT=300000 claude, or reconnect proofcut in /mcp
once the download has finished.
Any other MCP client runs the same server, from a checkout or with no checkout at all:
uv run --project /path/to/proofcut proofcut mcp
uvx proofcut mcp
On your own recording
uv run proofcut init myproject
uv run proofcut -C myproject import VO.wav --clip-id vo
uv run proofcut -C myproject transcribe vo # whisper, word-timed
uv run proofcut -C myproject seed vo # auto-editor strips silences
uv run proofcut -C myproject transcript vo --search "here's the thing"
uv run proofcut -C myproject cut vo 111:114 --plan # what do those indices say?
uv run proofcut -C myproject cut vo 111:114 --pad 0.1 # inclusive word range
uv run proofcut -C myproject export final.mp4 --render # or a .kdenlive to finish in an NLE
uv run proofcut -C myproject verify final.mp4 # did the render say what you edited?
To watch the edit instead, open the workspace:
uv run proofcut -C myproject open # server plus an app window
uv run proofcut -C myproject web --open # the same page in a browser tab
proofcut is 0.x software. A project from an older version is refused rather
than guessed at, and proofcut migrate brings it forward.
What it puts on your machine
proofcut is a local tool that needs real media binaries, so on a bare machine
proofcut setup downloads about 2.2 GB, most of it whisper. Here is
all of it, where it goes, and what takes it away. Sizes are a Linux x86_64
bare machine; proofcut setup --plan prints yours.
"Setup's folder" is one directory: ~/.local/share/proofcut/deps
($XDG_DATA_HOME if you set it), or %LOCALAPPDATA%\proofcut\deps on
Windows. It holds everything except whisper, which is a uv tool because that is
how whisper ships.
--planwrites nothing at all, and its exit code reports only what is missing (test_setup_plan_writes_nothing_and_exits_by_what_is_missing), so reading the plan can never turn into performing it.- Doctor's report is the input. A piece is installed only when its row is ✗. A tool you already have is never replaced, and never upgraded behind your back.
- Every download is pinned by URL and SHA-256, bumped by hand, never
resolved from a
latesttag. A hash that does not match leaves nothing behind. - No sudo, no administrator rights, no distribution packages, and nothing is ever written over an existing file.
- Every link, file, folder and uv tool is recorded, and
--uninstallremoves exactly those, including the Python uv fetched for whisper.tests/test_install.pyinstalls the lot into a fake home, uninstalls, and asserts the home's listing is what it was before (test_install_then_uninstall_leaves_the_home_as_it_was), so "removes exactly what it added" is checked on every run of the suite, not just meant. A link you repointed yourself is left alone, because it is yours now.
To see a removal before it happens:
proofcut setup --uninstall --plan
Two things setup deliberately does not do: it is a command you type and never an MCP tool, so an agent cannot start a 2 GB download or change your PATH; and it touches nothing outside your own user account.
Help wanted: a Mac or a Windows run
GitHub's macOS and Windows runners take the demo to a checked render, but a runner never reads the instructions. By hand, so far:
- Mac: one person, on an Intel Mac (i7-8850H, macOS 15.7.9). On Apple silicon only the runner has.
- Windows: the author's own Windows 11 laptop, twice, and nobody else's PC. Windows 10 and ARM64 PCs have not been tried at all.
If you have one of these machines and half an hour, one script installs what proofcut needs, makes a short test video, has proofcut cut, score, master and check it, and puts a report on your Desktop. It asks before it starts, records what it added, and removes exactly that on request, nothing you already had. A run that stops at the first step is just as useful, because where it stops is the finding.
git clone https://github.com/tydude001/proofcut
bash proofcut/scripts/mac_trial.sh
It downloads uv into ~/proofcut-mac-trial and uses it to run
proofcut setup, which fetches whatever of ffmpeg, auto-editor, Shotcut's
renderer and whisper your Mac is missing (§ What it puts on your
machine). It asks for no password; an Intel
Mac needs Apple's Command Line Tools, and the script says so if they are
missing. bash proofcut/scripts/mac_trial.sh --uninstall runs proofcut setup --uninstall and deletes that folder.
Then file the report.
git clone https://github.com/tydude001/proofcut
powershell -ExecutionPolicy Bypass -File proofcut\scripts\windows_trial.ps1
It downloads uv into one folder under %LOCALAPPDATA% and uses it to run
proofcut setup, which fetches whatever of ffmpeg, auto-editor, Shotcut's
renderer and whisper your PC is missing. Nothing is installed system-wide
and it needs no administrator rights. The same command with -Uninstall
runs proofcut setup --uninstall and deletes that folder. It puts
proofcut-windows-report.zip on your Desktop with your home folder's name
taken out; file the report.
Requirements
proofcut is developed on Linux (a Fedora-based desktop). On macOS and Windows the test suite passes on CI; where each stands by hand is § Help wanted above, and in detail docs/plans/PORTABILITY.md.
Every hard part of an editor already exists as mature open source; proofcut
is the layer that lets an agent drive those tools and check what they
produced. proofcut doctor checks everything below at once. On Linux,
Windows and either kind of Mac, proofcut setup installs whichever of the
last four doctor marks ✗: a static ffmpeg, whisper, auto-editor's release
binary and Shotcut's melt, which on Linux is the portable build that renders
with no display at all
(§ What it puts on your machine;
docs/plans/INSTALL.md).
Working on proofcut
CONTRIBUTING.md is where to start: where things live, how to run the checks, and the rules a pull request is checked against. A coding agent starts with CLAUDE.md, which Claude Code loads automatically and which holds the traps this repo has already hit. SECURITY.md says how to report a vulnerability.
The documentation:
- docs/MANUAL.md: every command, with the reasoning.
- docs/DEMO.md: the whole loop in two minutes.
proofcut's reasoning is part of what it ships, so the design record is public:
- PLAN.md: architecture, stack decisions, open questions.
- HISTORY.md: the dated record of what shipped and what the evidence said.
- COMPETITORS.md: what else exists in this space, and what proofcut does that they don't.
- NEXT.md: the directions after the queues closed, ranked.
- TRIAL.md: an agent cutting a video end to end, unattended and scored.
- docs/plans/: the plans, in progress and finished. A step that landed says "Shipped" and names its HISTORY.md section.
License
PolyForm Shield 1.0.0. proofcut is source-available, not open source: you can read, run, change and redistribute it for any purpose except building a product that competes with it. Cutting your own videos, running it for clients, building on it and forking it to fix a bug are all fine. For a commercial licence, ask.
The bundled typefaces are not proofcut's to relicense. The caption face in
src/proofcut/fonts/ and the three browser faces in src/proofcut/web/ are
OFL-1.1, each with its licence text beside it and its source in that
directory's FONTS.md.
Say thanks
If proofcut cut a video for you, you can buy me a coffee on Ko-fi.
