Open Source Agentic Video Editors

How These Were Verified

Every star count, fork count, licence and last-push date below was read from api.github.com/repos/{owner}/{name} on 4 August 2026, and every project's README was read before it was described. Nothing is listed that was not opened.

Three things follow from that, and they are worth stating because most listicles in this niche state none of them. Star counts move. Treat the numbers as a snapshot with a date attached, not a ranking — re-run the call yourself if the difference matters. Stars are not maturity. One project here has 467 stars and was last pushed on the day it was created; another has 99 stars and ships tagged releases to PyPI. And licence is a fact, not a vibe — GitHub's licence field was not trusted on its own here. Every repository that showed blank was opened and its LICENSE path, README and packaging metadata read by hand, which turned one blank into a declared ISC licence with a missing file. The reverse was checked too: one repository shows a licence GitHub names cleanly and is still not open source. Both cases are named further down.

Ordering is by star count descending, because that is how the search results order them and pretending otherwise would be a different kind of dishonesty. It is not a quality ranking.

Four Shapes, Not One Category

"Open source agentic video editor" covers four structurally different things, and picking the wrong shape wastes an afternoon before you find out. The question to ask first is where does the agent live, and what does it drive?

Shape 1 is the most popular right now for a simple reason: the hardest part of an agentic editor is the agent, and if the user already has Claude Code open, the project only has to supply skills and scripts. Shape 4 is the one to read the README about — the code being on GitHub does not mean the editing is happening on your machine or that it is free. We keep a longer comparison of video editing MCP servers for shapes 3 and 4 specifically.

The Roster

Ten verified projects, ordered by stars on 4 August 2026. Full write-ups follow.

The Projects

Also in the Field

These came out of the same GitHub search pass on 4 August 2026 and their numbers are from the same source, but they are listed rather than reviewed — this page only describes what its author actually read. Several are moving quickly enough to deserve a look.

The rate of arrival is itself the finding. Seven of those eleven were pushed within a week of this survey, and a category where a third of the notable repositories are under six months old is a category where the right answer changes between the time you read a comparison and the time you commit to one.

Read the Licence Before the README

Two of the repositories in this survey carry no licence statement anywhere — CutClaw (942 stars) and video-editing-mcp (284). No LICENSE file, and nothing in the README or the packaging metadata either; both were checked by hand on 4 August 2026, not merely read off GitHub's badge. A public repository with no licence is not open source. Copyright attaches automatically, the default is all rights reserved, and pushing code to a public host grants the world the right to view and fork within the platform's terms — not to use, modify or redistribute it. That is a fixable oversight and usually is one; the fix is an email to the author, not an assumption.

A third case is worth separating from those two, because lumping them together would be unfair to the author. vidpipe (207 stars) declares "license": "ISC" in its package.json and carries an ISC badge in the README — but the badge links to a ./LICENSE file that is not in the repository, so GitHub detects nothing. The intent is stated and npm consumers will read it; the grant itself has no text to point at. That is a missing file rather than a withheld permission, and it is a two-minute fix for whoever notices.

The trap in the other direction is a repository that has a licence, states it plainly, and still is not open source. Crayotter (187 stars) ships PolyForm Noncommercial 1.0.0 — a real, deliberate, well-drafted licence that permits use only for noncommercial purposes. It fails the Open Source Definition on field-of-use restriction, so it is source-available rather than open source, and if you are editing videos for a client or a company you have no grant. There is a LICENSE file, GitHub names it in the sidebar, and the whole thing is above board — which is exactly why it slips past a reader who has learned to treat a filled-in licence field as the all-clear. Read the name, not just the presence.

Among the ones that do carry a licence, the practical split is between permissive and copyleft. MIT and Apache-2.0 — video-use, VideoAgent, ave, video-audio-mcp, diffusionstudio/agent, kinocut, FireRed-OpenStoryline — let you build a commercial product on top and keep your changes to yourself; Apache-2.0 additionally grants patent rights, which is why it is the more conservative choice for a company. AGPL-3.0 — OpenMontage, OpenChatCut — is network copyleft: if you run a modified version as a service that users interact with over a network, you owe those users the source of your modifications. For personal use that is irrelevant. For a startup planning to host it, it is the whole decision.

There is a second layer worth checking that no licence covers: the assets and the model terms. FireRed-OpenStoryline says outright that its bundled fonts and music are deliberately basic because of licensing restrictions, and that the demo quality comes from a commercial asset library you have to supply. Any project generating images, voices or music is also passing your usage through a provider whose own terms decide what you may do commercially with the output. The repository licence covers the code. It does not cover the fonts, the tracks, or the model.

The Real Cost of Self-Hosting

The claim "open source is free" is true about the licence and false about the pipeline. Editing a video with an agent costs money at five distinct stages, and cloning a repository removes exactly one of them: the vendor's margin. Everything else moves onto your bill or onto your calendar.

Take those stages one at a time. Agent turns are usually the largest line and they do not change at all. The same tokens cost the same whether your script sends them or a vendor's does, and an agent that plans, calls tools, reads results and revises sends a great many of them. Anyone who has watched a long agent session burn through a rate limit knows this is not a rounding error.

Indexing — speech recognition, and in some architectures a vision pass over frames — scales with minutes of source, and this is the one place a GPU genuinely earns its keep. It is also where the idle problem bites: rented on-demand A100 capacity was tracked across providers in mid-2026 at roughly $1 to $5 per GPU-hour depending on who you rent from, and an instance you leave running bills at that rate whether it is transcribing or sitting still. If you own the card already and your electricity is somebody else's problem, this line goes to nearly zero and self-hosting starts winning early.

Rendering is H.264 encoding: linear in output minutes and resolution, mostly CPU, and boring. It is the most predictable cost in the pipeline and the one least worth optimising. Storage and egress are per gigabyte and video is large — a three-hour ProRes capture is not a text file, and the difference between keeping originals and keeping proxies is most of your bill. And maintenance is the item that sinks self-hosting projects that penciled out fine on the spreadsheet: ffmpeg filter graphs break on variable frame rates, rotation metadata, unusual pixel formats and audio streams with no timestamps, and every one of those is a real afternoon.

None of which makes self-hosting the wrong answer. It makes it a specific answer. Here is the honest split.

If you fall into the top four rows, stop reading vendor pages — including this one — and clone something. Privacy and control are not preferences that a lower price can outweigh, and a repository you can read is worth more than a feature you have to request. The rest of the table is where a hosted agent earns its margin, and it earns it on time and idle capacity rather than on capability.

What the Open-Source Field Is Still Weak At

Reading ten of these back to back, the gaps are consistent, and they are not the gaps you would guess.

• Most of them edit blind. Rendering a file and reporting success is not the same as looking at what came out. Only three — video-use, ave and CutClaw — have an explicit review step that inspects the rendered output and can send the plan back for another pass. Everywhere else, the first eyes on the result are yours.
• Nobody publishes a ceiling. Not one README in this survey states a maximum input length or file size. That is honest in a sense — a local tool's limit is your disk — but it means you find out where the pipeline falls over by feeding it a three-hour recording and watching.
• Long-form is barely addressed. CutClaw is the only project here explicitly built for hours of footage, and it approaches it through music rather than speech. Most of the rest assume clips or a short take, which is a different problem from a podcast.
• The agent gets shell access. Every shape-1 project works by letting a model run commands against your filesystem. That is the source of the power and it deserves a moment's thought about which directory you run it in.
• Setup is the real barrier, not capability. Between conda environments, Git LFS weights, Node runtimes for Remotion, four provider keys and a GPU driver, the distance from git clone to first rendered frame is usually measured in hours. The projects that solve this — a pip install and an ffmpeg binary — punch far above their star count.
• Abandonment is common and invisible. Two of the ten here were last touched over a year ago, and two were last touched on the day they were created. Stars accumulate; commits do not. Check the push date before the badge count.

The counterweight is that the good ones are genuinely good. video-use's decision to have the model read a packed transcript instead of frame-dumping — its README does the arithmetic at 30,000 frames against roughly 12KB of text — is a better piece of engineering than most of what is shipping commercially, and it is sitting there under an MIT licence.

Where Valmera Fits, and Where It Does Not

Valmera is hosted and closed-source. It does not belong on the list above and is not offered as one of the ten. It is here because it publishes this page and because it is the shape people usually land on after bouncing off a self-hosting attempt, and pretending otherwise would waste your time.

What it is: an agentic video editor in the browser. Upload real footage — up to 14 GB or 3 hours, MP4, MOV, MKV or WebM — describe the edit in plain English, and an agent indexes it (word-level transcript with speaker labels, silence detection, shot detection, and labeled frame tiles it actually looks at), edits an EDL, renders a preview from a fast proxy, looks at the frames it produced, and exports the final file from your original at source quality. Every reply is verified server-side against the edit decisions actually recorded, so it cannot claim an edit it did not make. The free plan is 50 one-time credits with no card.

The part that is relevant to this page: the interface is public. The same registry of 108 tools the in-house agent uses — 97 editing, 11 session — is published as a remote MCP server over streamable HTTP with OAuth 2.1, dynamic client registration and PKCE, and the tools are not re-declared for it, so there is no second list to drift. That makes it inspectable and swappable at the protocol boundary — you can point Claude at it exactly as you would point it at kinocut. It is not a substitute for a licence to read the code, and it would be dishonest to present it as one.

It is also not a text-to-video generator, which is what separates it from OpenMontage and from a good half of what is trending in this category. And it has real gaps: no multi-cam sync, no SRT or VTT (captions are burned in), no true crossfade, no audio denoise, no team seats, no direct publishing, and one deliverable per request rather than a batch. If any of those are the job, one of the projects above — or a traditional NLE — is the better answer.

Frequently Asked Questions

If You'd Rather Not Maintain a Pipeline

50 free credits, no card. Upload real footage, describe the edit, export from the original file. Or connect the same 108-tool registry to Claude over MCP.