What is OpenScreen?
OpenScreen is an MIT-licensed desktop app for Windows, macOS and Linux that records a window or the whole screen, with microphone, system audio and an optional webcam, and turns the take into a polished demo: zooms that follow the cursor, trims, speed ramps, text and arrow annotations, backgrounds and on-device captions. It exports MP4 or GIF, rendered on the GPU.
The app ships an MCP server in two forms: a stdio server started by running the app's executable with the mcp argument, and an optional in-app HTTP server bound to 127.0.0.1 behind a bearer token. The project is a community continuation of Siddharth Vaddem's original, which he archived after v1.5.0; the new repository has about 7,020 stars and released v2.2.2 on 4 November 2026.
How to install OpenScreen
claude mcp add openscreen -- "<executable>" mcpIs OpenScreen worth it?
OpenScreen gives an agent the whole loop from capture to finished video, not just an editing API. By our count of the source, the server exposes about 55 tools: listSources, startRecording and stopRecording for capture, getFrames so the agent can look at what it shot, exportVideo and exportAudio, an on-device Whisper transcribe tool, and some 40 editing tools shared with the app's own assistant, among them addTrim, addZoom, addAutoZooms, addAnnotation, removeWords and setCaptionSettings. The companion openscreen-demo skill strings these together: plan a storyboard, record the app, edit, export, check frames, then attach the MP4 to a pull request.
It can record any window or the full screen, the microphone, system audio, the webcam as a separate file, and the cursor's shape and clicks. That is the security point. A stdio server runs with your user rights and, per the docs, can record and edit without the MCP edits switch, so a prompt-injected agent could start a capture while a password manager or private chat is on screen. Recordings, transcripts and captions stay on disk, and OpenScreen runs no servers of its own; frames the agent pulls with getFrames, and transcripts, do go to whichever model provider your client uses.
The guard rails are better than most. The in-app HTTP server is off by default, listens only on 127.0.0.1 on port 47821 unless changed, checks a bearer token and the Host and Origin headers against DNS rebinding, and starts read-only until MCP edits is turned on. No tool deletes a project, and edits to a project not open in the editor are refused until the agent creates a checkpoint, of which the last 20 are kept in memory. The skill drafts an X post but, by its own instructions, posts only after you say yes.
The catch: the README calls the project under active development and warns of breaking changes to the .openscreen format and the CLI. The Windows .exe is unsigned, so SmartScreen warns on it; the Store build is signed. The app also sends an anonymous daily usage ping to PostHog by default, with no install ID, which can be turned off and is skipped when Do Not Track is set.
Our take: for a team that wants every UI pull request to carry a short video, this is the first tool to set up. Run the stdio server only in sessions where you would also let the agent take screenshots.
Where it earns its place
A front-end developer on a six-person SaaS team
Attach a 30-second video of each UI change to its pull request
Install the openscreen-demo skill, ask the agent to add a video of the change, and it records the local app, trims dead air, adds click zooms and attaches the MP4 with gh pr comment. Free; the model tokens are the only cost.
A developer-relations engineer at a dev-tools start-up
Turn a raw walkthrough recording into a launch clip with captions
Record by hand, then have the agent remove filler words with removeWords, add zooms and on-screen annotations, and export a 16:9 MP4 and a square GIF. The X post it drafts goes out only after approval.
A technical writer maintaining product docs
Re-shoot the same tutorial clips after each UI release
Keep a storyboard in the repo and let the agent re-record and re-export each clip through the CLI and MCP tools, then review the frames it returns before replacing the files.
Before you commit
- Setup
- Install the app (Microsoft Store, Homebrew cask, or a .deb, .rpm, .pacman or AppImage from Releases), grant Screen Recording and Accessibility on macOS, then claude mcp add openscreen -- "<executable>" mcp, or run npx skills add getopenscreen/openscreen --skill openscreen-demo. Codex, Cursor and plugin installs are documented.
- Integrations
- About 55 tools by our count of the source: capture (listSources, startRecording, stopRecording), inspection (getFrames, getCurrentDocument, getTranscript), editing (addTrim, addZoom, addAutoZooms, addAnnotation, addEffect, addAudio, setCaptionSettings), project tools (createProject, importMedia, setProjectInfo) and export (exportVideo, exportAudio). Over HTTP, write tools need MCP edits turned on.
- Data and privacy
- Recordings and projects stay on disk; transcription is on-device and OpenScreen has no servers. Frames and transcripts the agent reads go to your model provider. The stdio server can capture the screen, microphone, system audio and webcam with your user rights. The HTTP server binds 127.0.0.1 with a bearer token stored via Electron safeStorage. An anonymous daily usage ping to PostHog is on by default and can be disabled.
- Cost at scale
- Free. The cost is model tokens: getFrames returns images, so a frame-by-frame check of each export adds image input to every run.
- Limits
- Undo checkpoints live in memory, the last 20, and are lost when the app or server restarts. macOS 15 and later re-ask for screen-recording permission periodically. Wayland click capture needs extra input permissions. Project format and CLI may change.
What’s good
- Record, edit, check frames and export in one server
- In-app HTTP mode is local-only, token-gated and read-only by default
- On-device captions; no OpenScreen servers
- Free, MIT, all three desktop platforms
What’s not
- Stdio mode records and edits without the MCP edits switch
- Frames and transcripts reach your model provider
- Breaking changes to project format and CLI expected
- Unsigned Windows .exe outside the Store
Best for
Developers and small product teams who want an agent to film, cut and attach a demo video for each feature it ships.
OpenScreen alternatives
OpenScreen: common questions
- Is OpenScreen worth it?
- As complete a way as exists to let a coding agent film and edit its own feature demo, and unusually careful about what that agent may touch. We rate it 4.2 out of 5.
- How much does OpenScreen cost?
- Free and open source (MIT) for personal and commercial use; no paid tier. The optional built-in editing agent uses your own model API key.
- What are the downsides of OpenScreen?
- Stdio mode records and edits without the MCP edits switch; frames and transcripts reach your model provider; breaking changes to project format and CLI expected; unsigned Windows .exe outside the Store.
- Who is OpenScreen best for?
- Developers and small product teams who want an agent to film, cut and attach a demo video for each feature it ships.
- How much does OpenScreen cost for a team or in production?
- Free. The cost is model tokens: getFrames returns images, so a frame-by-frame check of each export adds image input to every run.
- Is OpenScreen safe for company data?
- Recordings and projects stay on disk; transcription is on-device and OpenScreen has no servers. Frames and transcripts the agent reads go to your model provider. The stdio server can capture the screen, microphone, system audio and webcam with your user rights. The HTTP server binds 127.0.0.1 with a bearer token stored via Electron safeStorage. An anonymous daily usage ping to PostHog is on by default and can be disabled.
- What are the best OpenScreen alternatives?
- The closest alternatives we have reviewed are OpenEdit, agentic-obs (OBS Studio MCP) and Talkthrough.
Answers checked Nov 11, 2026.
Make OpenScreen? Show this rating on your site
