Agent Mission Control

Keep long Claude Code runs on track, without sitting at the keyboard. Give the run a goal with a finish line, cap what it can spend, hold risky commands until you say GO, and steer it from the chat with /mc.

• A goal with a finish line. While a goal is active, a Stop hook keeps the run going until the agent marks it done (it's told to only when your "done when" holds). It lets the run stop when it's waiting on you, out of budget, or stuck, so it can't loop.
• Budgets with teeth. Cap tool calls, minutes and automatic continues. When one runs out, every tool call is refused and the agent is told to wrap up. A new goal starts a fresh budget.
• GO / NO-GO for risky commands. git push, npm publish, rm -rf, or any pattern you add, is held until you approve it. The agent works on other things or stops, then runs exactly the command you approved.
• Questions and notes mid-run. The agent asks with mc ask. You answer with /mc answer, and your notes reach it between tool calls. No restart.

One JSON file of state per project. Zero dependencies. Needs Node.js 18.17+ on your PATH. Tested on macOS with Claude Code 2.1.287; Linux and Windows aren't tested yet.

Install

In Claude Code:

/plugin marketplace add long-run-labs/agent-mission-control
/plugin install agent-mission-control@long-run-labs

The plugin does nothing in a project until you give it a goal there.

Use it

Start a mission in any project by typing, in the chat:

/mc goal Ship the CSV export --done-when tests pass and a PR is open

Then work as usual. Steer with /mc whenever you like:

/mc status                          goal, budget, open questions, held commands
/mc approve a1                      or: /mc deny a1 not on Fridays
/mc answer q2 Postgres
/mc say use the staging DB only     a note for the agent
/mc pause                           /mc resume
/mc set maxToolCalls 2000           /mc set maxMinutes 600
/mc set requireApproval git push, npm publish, terraform apply
/mc reset-budget                    /mc done
/mc help

When the agent is waiting on you (a held command or a blocking question), your /mc approve or /mc answer lets it carry on right away. A /mc typed while the agent is working waits until the run stops. To send it sooner, press Esc first: that stops the current step, and the agent picks up again after your /mc. Or use mc in a terminal (below), which acts at the agent's next tool call without interrupting it. If /mc clashes with another command you have, type /agent-mission-control:mc instead.

Three ways people use it

1. An overnight run. /mc goal Migrate the API tests to Vitest --done-when npm test passes, then /mc set maxMinutes 600 and /mc set maxToolCalls 2000. The agent keeps going until the tests pass or the budget runs out, and anything matching git push waits for you in the morning.
2. A risky refactor with a seatbelt. /mc set requireApproval git push, rm -rf, prisma migrate. The agent stops at each matching command and runs it only after /mc approve.
3. Steering without restarting. Mid-run, press Esc and type /mc say keep the old endpoint working until v3, or run mc say … in a terminal so nothing is interrupted. The agent gets the note and carries on, and /mc status shows what it's doing and what's next.

From a terminal

A note sent from a second terminal reaches the agent mid-run, after its next tool call. /mc help prints the exact command for your machine; from the project folder it looks like node "<plugin folder>/cli/mc.mjs" say "use the staging DB only". The plugin folder changes with each plugin update, so run /mc help again after one.

How it works

Hook event	What it does
SessionStart	Gives the agent the mission brief: goal, budget, open questions, recent log, protocol
UserPromptExpansion	Applies the /mc commands you type
UserPromptSubmit	Passes along anything new from you
PreToolUse	Refuses tools while paused or over budget; holds commands that need a GO; keeps the agent out of .mission/
PostToolUse, PostToolUseFailure	Counts the call; delivers your notes, answers and approval decisions
Stop	While a goal is active, keeps the run going until it's done, unless it's waiting on you, out of budget, or stuck

The agent reports with its own mc commands (now, log, ask, reply, done). It can't run yours: approve, resume, set, reset-budget and the rest are refused when it tries, and so is any command chained onto mc. A /mc only counts when a person types it in a session they're attending. The usual ways around that are refused: a nested claude -p "/mc …", a scheduled /mc, a script that carries one. And starting any nested claude mid-mission waits for your GO. Like the approval patterns, this is a seatbelt, not a sandbox: an agent that deliberately disguises a command can get past it. The skill tells it not to, and Claude Code's own permissions still apply. If the state file can't be read, tools are refused with the reason instead of the checks silently switching off.

Approval patterns match the whitespace-collapsed command, case-insensitively; * matches anything. They're a seatbelt, not a sandbox: Claude Code's own permission system still applies.

With Claude Code's /goal

Claude Code's built-in /goal keeps a session going until a condition holds. Mission Control has its own finish line, so you don't need both, and it adds what /goal doesn't: commands that wait for your GO while the agent keeps working, budgets that refuse tools once they're spent, and notes and answers mid-run. If you do run both, each keeps the run going until its own condition holds.

What it runs, stores and sends

• Runs: one Node.js script, hooks/mission-hook.mjs, on seven Claude Code hook events, and cli/mc.mjs when the agent or you run mc. Nothing else, and nothing is downloaded.
• Stores: .mission/mission.json in your project (goal, controls, status, questions, notes, approvals and a log) and a .gitignore beside it, so it stays out of commits. Nothing is stored anywhere else.
• Sends: nothing. This plugin makes no network connections and collects no data.
• Adds to Claude's context: the brief at session start, and your notes and decisions after tool calls, only in projects where you've set a mission; and the result of each /mc you type.

Remove a project's mission by deleting its .mission/ folder. Uninstall the plugin with /plugin uninstall agent-mission-control@long-run-labs.

Install without the plugin system

git clone https://github.com/long-run-labs/agent-mission-control
node agent-mission-control/install.mjs /path/to/your/project

This copies the runtime to <project>/.mission/ (git-ignored), registers the hook for seven events in <project>/.claude/settings.local.json (existing entries kept, original saved once as .bak, kept out of commits through .git/info/exclude), and adds the skills, /mc included. Add --shared to use the committed .claude/settings.json instead, and commit .mission/ too. In a project installed this way, that copy is in charge and the plugin stays out of the way.

Support

Questions, bugs and ideas: GitHub Issues. npm test runs the end-to-end checks against temporary projects.

The full kit

A paid kit with a live console and phone alerts used to be sold separately. It is no longer sold. This free plugin works on its own.

License

MIT