Vidmatic

Screen recording for AI coding agents: read a recording's transcript, marks and frames over MCP

View on vidmatic.ai

Hosted MCP Server

npx add-mcp 'https://api.vidmatic.ai/mcp'

Installs into Claude Code, Codex, Cursor and more

Documentation

Screen Recording MCP Server

Give your coding agent your screen. It reads what you said, what you circled and when, then files the work and sends back a narrated demo.

Server URL:

https://api.vidmatic.ai/mcp

Transport:

Streamable HTTP

Auth:

OAuth 2.1 browser sign-in, no key needed

Connect Claude Code

  1. 1. Run this in your terminal (once per machine) claude mcp add --transport http --scope user vidmatic https://api.vidmatic.ai/mcp Copy
  2. 2. Run /mcp, choose vidmatic, then choose Authenticate. Your browser opens: sign in and pick what the agent may do.
  3. 3. The tools appear in the same session. No restart.

Connect Cursor

  1. 1. Create ~/.cursor/mcp.json with this, or add the vidmatic entry to the mcpServers you already have {"mcpServers":{"vidmatic":{"url":"https://api.vidmatic.ai/mcp"}}} Copy
  2. 2. Open Cursor's MCP settings and start the sign-in for vidmatic. Your browser opens: sign in and pick what the agent may do.
  3. 3. The Vidmatic tools appear in Cursor's agent chat.

What your agent can do

  • /mcp__vidmatic__workstream_from_video Turn a screen recording into one GitHub issue grounded in your code, then work it and prove it.
  • /mcp__vidmatic__issue_intel Read a recording and post what the user actually asked for, R1 to Rn, as one issue comment.
  • /mcp__vidmatic__workstream Mirror your coding session live on Agent Workstreams, stage by stage.
  • /mcp__vidmatic__add_issue File one or more issues in your own tracker, up to ten per run, grounded in your code, each queued on its own workstream board.
  • /mcp__vidmatic__dispatcher One loop run every 10 minutes from the folder that holds your checkouts: it files queued change requests in the right repo, then hands the ready issues of each repo you enable to one orchestrator that ships them in parallel lanes.
  • /mcp__vidmatic__change_request_dispatcher The older change request watcher: now runs one dispatcher tick that only files queued requests, and points you at the dispatcher loop.
  • /mcp__vidmatic__auto-implement-beta Beta: dispatch each subagent with a brief and evals; it finishes only with evidence they pass. No issue URL works every open issue in the repo.
  • /mcp__vidmatic__auto-job-dispatcher The older one-repo factory loop: now runs one dispatcher tick that only ships ready issues in the repo you are in, and points you at the dispatcher loop.
  • /mcp__vidmatic__workstreams Nudge every stuck workstream: your agent checks each one and moves it forward or closes it.
  • /mcp__vidmatic__record Record and narrate a demo of the change you just shipped, delivered to your Inbox.
  • /mcp__vidmatic__before_after Record one half of a before and after pair for a bug fix.
  • /mcp__vidmatic__demo_user_journey A narrated demo of the shortest new-user journey to a first win.
  • /mcp__vidmatic__explain_implementation On request: a narrated explainer of how a shipped change was built, from the ask through the decisions to the code.
  • /mcp__vidmatic__audit_ux A narrated UX critique of a live page, each finding circled on screen.
  • /mcp__vidmatic__optimize Find the detours in finished coding sessions, on your own machine.

In Cursor, ask the agent to call get_skill(name="record") (any skill name above works the same way).

See it end to end: screen recording for AI coding agents, screen recording for Claude Code, screen recording to GitHub issue and AI agent proof-of-work demos.

OAuth 2.1, token lifetimes and where data lives: Vidmatic security

Tools, access levels and keys: MCP reference

Questions

What is the Vidmatic MCP server?

A hosted MCP server at api.vidmatic.ai/mcp. It lets your coding agent read your screen recordings (transcript, marks, frames), file work from them, and publish narrated demos.

Do I need an API key?

No. Register the server with one line or one config entry, then sign in from your browser when your client asks you to authenticate.

Which clients are supported?

Claude Code and Cursor. We list a client only when the sign-in redirect it uses is known for a fact (a recorded connect or the client's own docs) and passes our end-to-end sign-in test.

Where do the skills come from?

The server serves them. Nothing is copied into your repo, so you always run the current version.

Something not working? Troubleshooting