Playback's MCP server

Playback’s distinguishing feature is that everything you can do by hand in the app, an AI agent can do for you. Playback ships an MCP server — a small program called playback-mcp — that an agent connects to and uses to drive a live editing session: list windows, start a recording, drop a zoom on the moment that matters, add a caption, and export, all programmatically.

How it fits together

Agent (MCP client) ──stdio──▶ playback-mcp ──socket──▶ Playback.app
                              (bridge)              (applies the command live)

Your MCP client (Claude, Cursor, OpenCode, and so on) spawns playback-mcp and talks to it over stdio. playback-mcp forwards each request to the running Playback app over a local socket, where it’s applied to your project immediately — the same edits you’d make by clicking.

Before you connect a client

You need two things in place:

1. Playback, installed. The playback-mcp server ships with the app, so downloading Playback is all it takes — there’s nothing separate to build or install. The client configs on the next page point at the playback-mcp command.

2. Playback, running. The app must be open for an agent to control it. playback-mcp answers the initial handshake on its own, but the moment an agent calls a tool that changes your project, it needs the live app to apply it to — otherwise you’ll get a friendly “launch Playback” error.

Watching the agent work

While an agent is connected, a status pill in the header shows Agent activity — connected. Click it to open the Agent Activity panel, which lists each connected session (active or inactive) and a live feed of the recent commands the agent has run. There’s a Clear button to reset the feed.

You don’t have to hand the session over completely — you can keep editing while an agent is connected. Just know that, as the panel warns, your changes and theirs can step on each other, so it’s usually calmest to let one driver work at a time.

Next