MCP for AI agents

uterm-mcp is an MCP server, built on the official MCP SDK, that exposes the uterm control surface to any MCP-aware client (Claude Code, Codex, Cursor, Continue, Zed, …). Once registered, an LLM can drive uterm directly through tool calls.

The 28 tools come in three families: 11 session tools (list, create, connect, read, watch, annotate, set mode, fan-out), 10 hijack and control tools (acquire a lease, heartbeat it, send, step, release, worker input mode), and 7 graphical tools (screenshot, click, drag, type, key, and the graphical hijack pair).

What the agent can do

• Discover sessions and participants.
• Request hijack subject to RBAC — the human operator can grant, deny, or share.
• Capture snapshots of the current screen, returned as text for the LLM to reason over.
• Place annotations at (row, col) so the human sees what the agent is pointing at.
• Participate in DeckMux chat — coordinate with humans in the same room, not in a separate channel.

Why this matters

A screen-scraping agent has to infer the terminal’s state from raw bytes. A uterm MCP agent sees the same emulator snapshot the browser sees, with structured presence and chat context. The agent is participating in the session, not eavesdropping on it.

Authorization is still enforced

The agent obeys the same hijack leases and role gates as every other participant. The MCP surface does not bypass the control plane — it is a client of it.