Unterm

Unterm

Terminal that AI agents can drive

Description

Running Claude Code, Codex and Gemini CLI at once usually means flipping through tabs to see which finished, which is waiting on you and which broke something. Unterm doesn't bolt a chat box into the terminal. It does the opposite and makes the terminal itself infrastructure agents can call: every window starts a local MCP server, and outside agents use 152 methods to spawn shells, run commands, read pane state, take screenshots and record sessions.

On the other side it is a cockpit for you: live agent state per pane, an Inbox that surfaces the agent waiting on you so Enter jumps straight to it, fleets of N agents on one task in isolated git worktrees, and a Review page to diff, merge or roll back. macOS, Linux and Windows, local-first and free.

Features



MCP control surface: 152 authenticated methods; Claude Code, Codex, Cursor, Aider or your scripts share one JSON-RPC API, bound to 127.0.0.1 with no cloud round trip.

Agent Cockpit: per-pane agent state with a waiting-first Inbox.

Worktree fleets: send one task to several agents, each in its own isolated git worktree.

Review and rollback: diff the results, merge what works and roll back what doesn't.

Governance: published risk levels per capability, expiring non-replayable leases, isolated workspaces and verifiable evidence bundles.

Web settings: theme, language and config edited in a browser page with live preview; nine UI languages.