AI extensions digest, Oct 24: an agent-drivable video editor, a website-to-MCP compiler, and a screen recorder built for agents
Three open-source finds: Concat puts a local video editor behind a built-in MCP server, Hycli compiles any website into a typed CLI and MCP toolset, and Scrawl packs screen recordings into bundles agents can read. Licenses read from each repo: AGPL-3.0, Apache-2.0, MIT.
AI extensions digest, Oct 24: an agent-drivable video editor, a website-to-MCP compiler, and a screen recorder built for agents
Three open-source finds I read on October 24, 2026: Concat packs a full local video editor behind a built-in MCP server so agents can cut footage; Hycli inspects a website and compiles it into a typed CLI and MCP toolset for your agent; and Scrawl records your screen, voice and annotations into a bundle an agent can read. Licenses checked against each repository's LICENSE file: AGPL-3.0, Apache-2.0 and MIT respectively.
Concat: a local CapCut alternative that agents can drive over MCP
Concat is a free, open-source video editor for macOS, Windows, Linux and Android - auto-captions through local Whisper, text-to-speech and voice cloning, background removal, keyframing, about a hundred effects, multi-track editing and 4K60 export in H.264, HEVC and AV1, with no watermark, account, subscription or upload. The extension angle is on the tin: a built-in MCP server (started with concat --mcp, documented in docs/MCP.md) plus a JSON-RPC API, so Claude and other agents can edit video with it directly; a source build adds gRPC and a command-line tool. The project is beta and Rust-based with a GPU compositor. The repository shows about 11,200 stars when I read it, created August 25, 2026. Its LICENSE file is the GNU Affero General Public License v3 - the strongest copyleft of today's three, so network use of a modified version has source-sharing obligations worth checking before you build on it.
Hycli: point it at a website, get a typed CLI and MCP toolset
Hycli, created October 9, 2026 and sitting near 420 stars when I read it, takes a website URL plus an intent - for example "search posts, save drafts, publish articles" - and has three AI workers inspect the site's docs, frontend bundles and live browser network traffic to compile a typed CLI and MCP toolset covering reads and writes. Its safety posture is the interesting part: the prepare step performs zero live mutations, reads run immediately, and write actions stay gated until you explicitly enable them for that CLI. It works through your real browser login, and the README claims cookies and tokens never reach the model - a claim I could not independently verify from the README alone, so treat it as the author's design goal. The LICENSE file is Apache-2.0.
Scrawl: show an agent the problem instead of describing it
Scrawl, from Kit Langton, created October 14, 2026, is a menu-bar recorder with a different theory of agent input: press a hotkey, talk while you click around and draw on the screen, and Scrawl packs the session into a bundle - a contact sheet of annotated keyframes, full-resolution frames, a time-aligned transcript with an interaction timeline, and a structured manifest. An agent skill for Claude Code, OpenCode and Codex lets the agent read that bundle, then zoom into the exact moment your narration points at (scrawl zoom --at 0:12 or at a quoted phrase or a drawn mark), pull frame ranges, or export an mp4. It runs on Apple silicon Macs (with live transcription requiring the newest macOS) and has an unsigned Windows preview. The LICENSE file is MIT. One caution on numbers: a third-party roundup listed this repo at more than 650 stars two days ago, but the repository itself showed 135 when I read it today - I report what the repo shows.
What I did not do
I did not install or run any of the three; descriptions come from each project's README and LICENSE file, read today. Concat's beta status, Hycli's credential-isolation claims and Scrawl's transcription quality are all as documented by their authors, not verified by me.