Agent Mission Control

One screen for every Claude Code session on your machine.

If you run four or five Claude Code sessions at once, questions and permission prompts get buried in scrollback and sessions sit blocked without you noticing. Mission Control puts each session on a station of a retro control-room floor. Anything that needs you arrives there as a card, with enough context to decide, and your answer goes straight back into the waiting session. The longer a station waits, the louder it gets.

It is local-first. The hub is a single Bun process on 127.0.0.1:4242 with SQLite in ~/.agent-mission-control/. Nothing leaves your machine unless you turn on the optional ElevenLabs voice.

Unofficial project

Agent Mission Control is an independent project. It is not affiliated with, endorsed by or supported by Anthropic. It relies on some Claude Code behaviour that is not documented and may change in any release: the per-session messaging socket, CLAUDE_CODE_SESSION_ID in the MCP server's environment, and some hook payload fields. It is alpha software: the current release is v0.1.0, and 0.x releases may change behaviour between versions.

If a Claude Code update breaks something, please open a bug with the output of claude --version and amc doctor.

Contributing · Code of conduct · Security · MIT licence

Related MCP server: cware-hil-mcp

What lands on the floor

The terminal prompt still appears alongside every card. Answer wherever you are looking: the first answer wins, and the card clears itself if you answered in the terminal.

When a turn ends with questions in prose, a Stop hook first sends the turn back once and asks Claude to use the question tool instead, so most of those become proper question cards.

What else is on a station

• Persona. Every session gets a callsign, a pixel sprite and a voice that lasts across resumes of that session.

• Crew bay. Subagents, in-process teammates and nested claude runs appear as mini sprites on the session that spawned them, not as stations of their own. Their questions land on the parent's station, tagged with the child's role.

• History. The last 80 tool calls as one line each, with what they touched, for example Edit · apps/web/src/hub/state.ts. It survives reloads and hub restarts.

• Context. Every card shows your last prompt and what Claude said just before asking, read from the session transcript.

• Escalation. A waiting station glows amber at 2 minutes, strobes red at 5 and sounds the klaxon at 10. An idle session, one that has finished its turn, shows a calm blue READY state that never escalates.

• Transmit box. Send any session a new topic or a comment. An idle session starts a new turn with it. A working one folds it into its current turn after the next tool call.

Requirements

• macOS or Linux (arm64 or x64). Messaging a session uses its local Unix socket.

• Claude Code 2.1.280 or later. The integration points were verified against that version.

• Nothing else for the amc binary. Building from source needs Bun 1.3 or later and go-task.

Quick start

Everything ships as one amc binary: the hub with its UI built in, the wiring commands and the MCP server Claude Code launches.

brew install dannyhawkins/tap/amc
amc start                       # the hub and its UI on :4242, in this terminal
amc wire --global --gate        # in another terminal: every Claude Code session reports in
open http://127.0.0.1:4242
claude -c                       # restart each running session once so it loads the wiring

Prefer a gentler start? amc wire --global --gate-questions-only routes only Claude's questions and plan approvals to the floor and leaves tool permission prompts in the terminal. Upgrade later with brew upgrade amc, then restart the hub when no cards are waiting.

No Homebrew? Download the tarball for your platform from GitHub Releases and check it against SHA256SUMS. docs/install.md has the steps for each platform, upgrading and uninstalling. amc doctor says what is wrong when something is.

amc wire --global merges hooks into ~/.claude/settings.json and registers the MCP server for all your projects as <path to amc> mcp, after writing a timestamped backup of the settings file. It keeps your existing hooks and settings. Wiring records where amc lives, so rewire if you move the binary (Homebrew's path stays put across upgrades). Running sessions need a restart (claude -c resumes the conversation) because each session loads its MCP servers when it starts.

The hub runs in the foreground of whatever terminal starts it; there is no login item yet (#1). After a reboot, run amc start again. Its data lives in ~/.agent-mission-control (AMC_DATA_DIR).

Upgrading from a checkout install. Wiring made by task wire* before amc existed points Claude Code at bun <checkout>/packages/mcp-server/src/index.ts. Run amc wire --global --gate (add --telemetry if you had it) and it rewrites every such entry it finds, user scope, local scope and each known project's .mcp.json, to the binary, and prints what changed. amc doctor lists any it missed.

From source

task install                    # bun install + git hooks
task start                      # build the UI and run the hub on :4242
task wire:global -- --gate      # same as amc wire, run from this checkout
task build:bin                  # or build dist/amc and put it on your PATH

From a checkout the MCP server is wired as bun <checkout>/apps/cli/src/main.ts mcp, so no build is needed. bun apps/cli/src/main.ts <command> is amc <command>.

Tear-down

To pause without unwiring, stop the hub with Ctrl+C. Sessions carry on as normal: hooks fail silently, gated prompts fall back to the terminal, and request_decision tells Claude to ask in chat instead.

To remove Mission Control completely:

# 1. Remove the global wiring: our hooks, env keys (including telemetry) and the
#    user-scope MCP server. Your own hooks and settings stay; a backup is written first.
amc unwire --global

# 2. Remove any per-project wiring you added, shared or --local.
amc unwire ~/Code/some-project

# 3. Stop the hub (Ctrl+C in its terminal), then delete its data: the SQLite
#    database with sessions, decisions and history, plus config.json.
rm -rf ~/.agent-mission-control

# 4. Optional: delete the settings backups that wiring and unwiring wrote.
rm ~/.claude/settings.json.amc-backup-*

# 5. Remove the binary.
brew uninstall amc              # or delete the amc you copied onto your PATH

Restart running Claude Code sessions afterwards so they drop the hooks and the MCP server.

Wiring options

Add --dry-run to amc wire to print what it would change, --port when the hub is not on 4242, and --home <dir> to wire a Claude config dir other than ~/.claude. From a checkout, task wire DIR=<dir> prints and task wire DIR=<dir> -- --write applies, as before.

Configuration

Set these as environment variables when starting the hub. ignoreCwd and socketReply can also go in ~/.agent-mission-control/config.json; environment variables win.

Using the floor

The mission log on the right lists every decision and alert, newest first. The CREW toggle shows or hides lines from child agents.

Voice

VOICE in the header reads out new cards, stations standing by, stations waiting five minutes and new sessions, in each persona's own voice. The caret next to it picks which of those are spoken, the rate, and the provider:

• Browser (default). The browser's own speech voices, offline. On macOS, download a Premium or Enhanced voice (System Settings, Accessibility, Spoken Content, System voice, Manage Voices) and restart the browser; stations prefer Premium, then Enhanced voices. Every station on the floor gets a different voice while there are enough, plus its own pitch and rate.

• ElevenLabs (optional). The hub renders each line with ElevenLabs and caches the mp3, so every line of every persona is paid for once. Each live session gets its own stock voice, the best fit for its character not already used by another live session (a pirate gets a husky trickster, a bureaucrat a mature, reassuring one), with delivery nudged per persona so no two sound alike. The popover lists each station's voice with a reroll button; a voice sticks to its session across resumes and hub restarts. When a line cannot be fetched (hub down, no key, daily cap reached, ElevenLabs error or timeout) that line is spoken by the browser voice and the popover says why.

Each persona always phrases an event the same way. When a session first speaks through ElevenLabs, the hub pre-renders its six core lines (on the floor, standing by, has a question, permission to run Bash, plan approval, waiting five minutes; about 250 characters); anything else is rendered the first time it is said.

To turn ElevenLabs on, give the hub a key. The recommended place is a file only you can read, which the hub re-reads when you open the voice popover, so a new or rotated key needs no restart:

printf 'ELEVENLABS_API_KEY=%s\n' 'sk_...' > ~/.agent-mission-control/secrets.env
chmod 600 ~/.agent-mission-control/secrets.env

The hub ignores the file, with a warning in its log, while anyone but you can read it. ELEVENLABS_API_KEY in the hub's environment works too and wins over the file. Watch for this when a key is exported in a shell profile (~/.zshrc): the hub silently uses that one and your secrets.env is ignored (the hub logs "ELEVENLABS_API_KEY from the environment wins over secrets.env"). Start the hub with env -u ELEVENLABS_API_KEY amc start to use the file. amc doctor reports which source the running hub uses and warns when your shell exports a key alongside secrets.env.

What is sent to ElevenLabs: only the fixed announcement lines, built from the callsign, a count, a tool or agent role name and the persona's stock phrasing ("Ahoy! Nova has a question."). Card content, prompts, questions and answers are never sent. The hub counts the characters it sends per day against AMC_TTS_DAILY_CHARS and keeps at most 50 MB of audio in ~/.agent-mission-control/voice-cache/, dropping the least recently played first.

Worth knowing

• Messages you send are relayed, not typed. Claude Code labels anything arriving on a session's socket as a message from another session. Claude acts on it for ordinary requests, but may confirm risky steps in the terminal. Sessions in bypass-permissions mode hold these messages until approved.

• The socket interface is undocumented. Replies and messages use Claude Code's cross-session socket, which a Claude Code update could change. When delivery fails, the card falls back to "answer in the terminal".

• Restarting the hub drops pending gate cards. Each gated prompt holds a connection open while it waits, so a restart sends those sessions back to their terminal prompt. Check the floor for pending cards before restarting.

• Headless claude -p runs have no question tool, so only permission prompts and request_decision apply to them.

Development

task dev            # hub with --watch, plus Vite on 127.0.0.1:5173 proxying to the hub
task sim            # three simulated sessions against a running hub
task check          # biome + tsc + bun test
task build:bin      # dist/amc for this platform, UI embedded
task build:bin:all  # dist/amc-<os>-<arch> for darwin and linux, arm64 and x64

Open http://127.0.0.1:5173/?mock=1 to run the whole UI against a built-in simulator, with no hub or real sessions needed. CONTRIBUTING.md covers setup, checks, commit messages and pull requests.

Roadmap and ideas

What is being considered next, tracked as issues. Comment on the ones you care about, or start a thread in Discussions.

• Start the hub on login, once it has proven stable in daily use (#1).

• A Claude Code plugin as a one-step install channel (#3).

• Custom per-persona voices with ElevenLabs Voice Design (#5).

• Hub-side voice with macOS say, so announcements work with the browser closed (#2).

• An npm / bunx distribution (#4).

Bugs and feature requests: open an issue.

Further reading

• Install: Homebrew, release tarballs and building from source, upgrading and uninstalling.

• Architecture: input channels, how answers are routed, and the trade-offs behind each route.

• Runbook: wiring details, what a card shows, and troubleshooting.

• Releasing: tagging a release from main, what the release workflow builds, and the Homebrew tap token.

• CLAUDE.md: verified Claude Code facts and design rulings, for agents working on this repo.

Licence

MIT. Contributions are welcome: see CONTRIBUTING.md and the code of conduct. Report security problems privately as described in SECURITY.md.