Unterm v0.55: the Agent Cockpit
Your terminal now sees every AI agent running inside it — who's working, who's blocked on you, and what they changed. Plus fleets: one task, three agents, three isolated worktrees, side-by-side review.
2026-07-11T00:00:00.000Z
If you run AI coding agents all day, you know the actual bottleneck isn’t the model. It’s you, alt-tabbing across panes to find out which agent finished, which one has been silently waiting twenty minutes for a y, and what exactly any of them did to your repo.
Unterm v0.55 makes the terminal answer those questions itself. We call it the Agent Cockpit.
Who’s doing what, at a glance
Every pane that hosts an agent — Claude Code, Codex CLI, Gemini CLI, Aider, and friends — now carries a live state: working (breathing blue dot on the tab), needs you (amber), done (green), idle (grey). The top bar keeps a cross-window tally that flips amber the instant any agent blocks on you.
Detection is zero-config for Claude Code and Gemini CLI: Unterm reads the signals agents already emit — title spinners, OSC progress sequences, desktop-notification escapes. One command, unterm-cli agent enable-hooks, upgrades Claude Code, Codex, and Aider to exact hook-level reporting (merge-only config writes, .unterm-bak backups, your existing settings untouched).
The Inbox: stop hunting for blocked agents
Ctrl+Shift+A opens the Agent Inbox: every agent across every window, sorted waiting-first, with how long it’s been stuck and what it’s working on. Enter jumps to the pane. The most common multi-agent action — “go press y somewhere” — is now two keystrokes.
Fleet: race three agents on the same task
unterm-cli fleet launch --agents claude,claude,codex -- fix the login bug
Each member gets its own git worktree (beside your repo, on its own branch — your working copy is never touched), its own tab, and its own state badge. Crew presets come from whatever agents you actually have installed, including claude + codex + gemini for a proper three-model bake-off.
Review: nothing an agent does is untracked
The moment an agent starts working in a repo, Unterm snapshots the worktree as a dangling commit — HEAD, index, and files untouched, so you’ll never notice it happened until you need it. The new Review page shows line-level diffs (including files the agent created), squash-merges the take you like into your repo as staged changes — the commit stays yours — discards the takes you don’t, rolls any repo back to any checkpoint, and puts two fleet members’ diffs side by side.
Also in this release
- CJK input methods now work correctly in every palette — composition previews inline instead of vanishing (this one was found by our own daily dogfooding, ten minutes after shipping the palettes).
- OSC 9 / OSC 777 notifications now actually reach window-level consumers.
- Sidebar no longer rebuilds ~44ms per title-spinner frame while an agent works — a real CPU saving for anyone running agents for hours.
The thesis, completed
Unterm’s founding idea is “the terminal AI agents can drive” — every window is an MCP server that external agents operate. The cockpit is the same idea pointed inward: the agents running in your panes are now first-class citizens the terminal understands. And true to that thesis, the cockpit itself is fully scriptable — 12 new MCP methods and three new CLI command families mean an orchestrating agent can launch fleets and review diffs without a human in the chair.
Free, MIT, local-first, no accounts, no telemetry. Download v0.55 · Agent Cockpit docs