autodemo

A screen recorder for AI agents. Tell Claude Code, Codex, Cursor or any MCP agent to "record a demo of the signup flow". The agent opens the app with the tools it already has (Claude in Chrome, computer use, a browser MCP) and drives it. autodemo films the tab or window and edits the result into a polished video:

• zoom on every action

• a smooth cursor and click ripples

• a local AI voice reading the agent's narration, with word-by-word captions

• macOS desktop or gradient backgrounds, with a window frame

• the agent's thinking time cut out

• a Loom-style share page

agent picks a tab/window → camera rolls → agent drives the app & narrates → camera_stop → render → agent reviews → studio

Why it works

• Like sharing a tab in Zoom, with no picker. The agent names a tab ("todomvc.com") or an app window. autodemo finds it in Chrome, Brave, Edge or Safari, brings it forward and films only that page. Browser bars and banners can be trimmed off.

• Uses the agent's own hands. The agent drives with its own browser or computer-use tools, logged in as you. If it has none, camera_act clicks and types for it.

• Any app or device. Web apps, native Mac apps, the iOS Simulator, Android emulators, iPhone Mirroring.

• Long takes. 10 to 20 minute walkthroughs, with chapters. If macOS drops the stream mid-take, it keeps recording in a new segment.

• Cheap on tokens. One call per step, one-word replies, small screenshots only when needed.

• Local and private. ScreenCaptureKit, Kokoro-82M voice (falls back to macOS say), ffmpeg. No cloud.

Related MCP server: screencast

Install

As a Claude Code plugin:

claude plugin marketplace add mdayan8/autodemo
claude plugin install autodemo@autodemo

Or as an MCP server for any agent:

claude mcp add --scope user autodemo -- npx -y autodemo mcp
codex mcp add autodemo -- npx -y autodemo mcp

Then copy or link skills/autodemo into your agent's skills folder: it teaches the agent to behave like a presenter.

The first time, macOS asks for Screen Recording permission for the app running your agent. Allow it, and Accessibility too if you want camera_act. Run npx autodemo doctor to check ffmpeg, the voice model and Chromium.

Then ask: "Record a 30 second demo of creating an invoice in my app, the tab with localhost:3000."

Looks

background: macos (desktop with menu bar and Dock), macos-light, aurora, grid, midnight, sunset, ocean, forest, peach, mono, light, any two colors, or your own image. frame: mac (traffic lights) or browser (address bar). output: 4K or 1080p, plus a vertical cut for Shorts and Reels.

Studio

autodemo serve (or the studio tool) opens a local page listing every demo your agents recorded, with live status while one is recording or rendering. Each demo opens its share page: player, chapters, transcript, captions and a call-to-action button.

MCP tools

CLI

autodemo serve                         # studio at http://localhost:4747
autodemo render demos/x                # re-render (new background, trim, captions) without re-recording
autodemo run demo.json --out demos/x   # scripted mode: record + narrate + render
autodemo doctor

Each demo folder contains:

• final.mp4 (and vertical.mp4)

• share.html

• captions.vtt

• transcript.md

• poster.jpg

• preview.gif

• timeline.json

• the raw segment-*.mp4 and recording.json, kept so you can re-render later

Limits

• Camera mode needs macOS 14+, because it uses ScreenCaptureKit. The Swift helper builds on first use and needs Xcode Command Line Tools. Scripted mode runs anywhere Playwright runs.

• The camera films the window's area of the screen. A cover guard pauses whenever another window or a notification is on top of it, and drops the moment before it noticed, so you can keep using your Mac during a take. Don't use the filmed window yourself while it records.

• macOS sometimes stops a capture stream on its own (about once a minute on some machines). autodemo starts a new segment immediately, losing about a second each time.