OpenEdit

An agent skill from VEED that turns a coding agent into a video editor: no GUI, no timeline, npx skills add veedstudio/open-edit and then “Add subtitles to my video [VIDEO]” in claude-code, Codex or Gemini. 209 stars, 14 forks, TypeScript, Apache-2.0 at ingest. Its tagline is “Not the editor you rent, but the one you own” — written by a company whose business is renting a browser-based video editor.

Tier T1 (official project repo, per the ladder), promotional and unmeasured. The one number in the README carries the README’s own disclaimer, which is unusual enough to quote below.

The second video-editing skill in this spoke, and it disagrees with the first

video-use arrived four days earlier from a different org, does the same job, and answers the central design question the other way.

The interesting divergence is the authoring surface. Both read a transcript, because per-word timings are what caption timing needs. But video-use then reasons in text all the way to the cut, while OpenEdit has the agent write HTML and CSS for the caption design — a language the model is fluent in, rendered by a binary that never launches a browser. text-first-agent-design says to hand the agent the representation the platform already maintains; this is the neighbouring move, handing it the representation the model already knows, which is the same instinct paperclip applies to scientific databases. “Move the text up a bit” and “make the ‘now’ really stand out” are CSS edits, and a reference image can be handed over as the style spec.

The precondition is enforced, not requested

Every transcription provider writes the same runs/<key>/transcript.json, and the README is blunt that “nothing downstream can tell which one ran” — provider choice is asked once, stored in .open-edit-prefs.json, and never surfaces again. On top of that sits a hard precondition: per-word timings are required whichever provider you use, because without them caption reveals drift out of sync, “so a transcript that has none is refused rather than rendered badly.”

That is a mechanical gate on an input, not an instruction to the model, and it belongs next to claude-security-plugin‘s Python-computed quorum in the small set of places where this corpus finds enforcement rather than prose (agent-guardrails). Cheap version of the same idea: check the precondition your quality depends on, and fail loudly.

What it says about itself

• The benchmark comes with its own caveat. “Internal benchmarks measured the renderer up to 2.2x faster than Chrome-driven renderers. These measurements are preliminary and not yet reproducible outside VEED; a documented benchmark and its methodology will follow.” A vendor number that names itself unreproducible is the published-limitations posture in a smaller space.
• Scope is bounded. “V1 targets captions. Motion graphics, charts, and brandbook-matched styling render today, but are less exercised than captions and should be expected to have rough edges.”
• The renderer is optional. Composition does not depend on it; Chrome can be the backend instead, which keeps the closed binary from being load-bearing.
• Platform lock is severe. Apple Silicon and macOS Tahoe 26.0; Intel Macs and earlier macOS are “not supported”; Windows and Linux are “planned; prioritisation depends on demand.”

What “open” covers here

Apache-2.0 for the editor, PolyForm Shield 1.0.0 for the renderer binaries — which permits commercial use of videos you produce with no payment to VEED, and does not open the renderer’s source. So the orchestration is open and the speed is not, with an escape hatch (Chrome) that costs you the 2.2×. Set it beside the spoke’s other “open” claims — openclaw‘s gateway, qm‘s MIT core with a hosted service, foundation-agents‘s community clone — and the recurring pattern is that the agent layer is what gets opened, because that is the part that spreads.

Composition with paid services is assumed rather than hidden: the README’s own examples call Seedance 2.0 and VEED Lipsync 2.0 on Fal for generation and lipsync, and Figma MCP to read a brand book before styling campaign cards.

The vendor shape

VEED sells a subscription video editor. This repo hands the same work to a coding agent the customer already pays for, and says so in the tagline. Read against the synthesis vendor ladder, it is a different move from Anthropic’s or Google’s: not a model vendor climbing into the harness, but an application vendor publishing the agent-native version of its own product, presumably betting that being the skill people install beats being the tab they cancel. The transcription funnel is where the business sits — VEED is the default provider and a free account covers about 10 minutes a month.

Related

video-use · veed · text-first-agent-design · agent-skills · agentskills-spec · agents-md · claude-code · agent-guardrails · watch-skill · paperclip · model-context-protocol · qm