Screen Recorder MCP

Screen Recorder MCP

Let your AI agent record a real browser page and cut it into a publishable vertical video.

By xukunAI · an MCP server for Claude Code, Cursor, Codex and Hermes. Your agent records a page over CDP (no browser extension, no permission dialog), then composes the clip: device frame, held-frame callouts, punch-ins, captions, music with real ducking, end card.

Everything runs locally. No upload, no API key, no cloud render.

Two recordings, one canvas

split puts a second recording next to the first — stacked in 9:16, side by side in 16:9, with ratio and gap to tune it. The direction follows the canvas, so the agent never has to choose. Captions, callouts, end card and music still apply in the same call.

{ "target": "vertical",

"split": { "source": "C:/rec/second.webm", "ratio": 0.5, "gap": 6 }, "captions": [{ "start": 0.8, "end": 5.0, "text": "before / after" }] }

Full guides: [English](docs/USAGE.md) · [中文](docs/USAGE.zh.md)

录制 → 设备框 → 冻结圈注 → 局部放大 → 字幕 → 结尾卡 → 成片

Why this exists

Every "AI video" tool generates pixels and hopes the text survives. This one doesn't generate anything: it records your actual UI and draws text with a font engine. Every digit is your real number.

The agent-facing part is the point: an agent can produce a finished clip as a tool call, not a chat suggestion.

Tools

compose_clip takes captions, annotations (held-frame callouts), focus (punch-ins), end_title, music, frame (phone/browser), pip (picture-in-picture), split (second recording on the same canvas).

Install

npx (recommended) — works in Claude Code, Cursor, Codex, Hermes and any MCP client:

{

"mcpServers": { "screen-recorder": { "command": "npx", "args": ["-y", "screenrec-agent"] } } }

Then verify the install:

npx -y screenrec-agent --selftest

First run downloads the agent (~33 MB) and a pinned ffmpeg build (~201 MB) into %LOCALAPPDATA%\screenrec-agent\bin. The cache is reused afterwards, and upgrading the launcher re-downloads only the agent — ffmpeg is not fetched again. Downloads resume and retry on a dropped connection instead of failing silently.

Or use the exe directly. Download screenrec-agent.exe from the latest release, put ffmpeg.exe beside it (only compose_clip needs ffmpeg), and point command at the exe.

Published in the official MCP Registry as io.github.Robert0303-C/screenrec-agent.

Pricing

The trial starts on your first successful compose, not on install. When it ends, compose_clip returns license_required with a purchase link — recording keeps working, nothing is deleted, nothing crashes. Check what is left with screenrec-agent --status, unlock with screenrec-agent --activate .

The only network call is activation itself. Trial state lives on your machine (three copies, clock rollback detection) because the alternative is a server that watches you use it. Deleting all three copies is equivalent to a fresh install — offline software cannot prevent that, and we would rather be honest about it than ship surveillance.

Buy: https://www.creem.io/product/prod_57FG0V8N2N4z3PrCzK0Klr · xukun.online

The one gesture Chrome insists on

chrome.tabCapture needs activeTab, and activeTab only comes from invoking the extension on that tab. No script can fake it (verified: even a trusted CDP click inside an extension page fails). So:

• Page recording (CDP) — fully autonomous, no click, no extension
• Desktop + microphone recording (extension) — the human clicks the toolbar icon once, the agent handles everything after

Requirements

• Windows 10/11 x64
• ~300 MB disk
• An MCP client

Roadmap

• CDP recording, freeze-frame callouts, punch-ins
• Device frame, captions, end card, music ducking, picture-in-picture
• Voice-over (piper, EN + ZH) — Full edition
• macOS

License

Commercial. Source available for review; redistribution of the packaged binary is not permitted.