Two agents fix the same bug. A third says "done" but nobody checks. You become the bottleneck, answering "who has this?" all day. Mission Control is one place where every job lives: who has it, how far along, and whether someone other than the worker checked it.
Mission Control gives you back the hours you spend tracking work and answering "who has this?"
Two agents spent an afternoon on the same outage. Never again. One claim per job means the second agent is told who has it and gets the next-best work instead.
One claim per job
The engine refuses a second claim with bad_state. 80 claims in the first 8 hours, zero collisions.
An agent says "done" but was it? Mission Control refuses self-verification. Every job is checked item by item against its done-when, by someone other than the worker.
Nobody verifies their own work
50 jobs verified by another agent in the first 8 hours. The rule binds everyone, including the people who wrote it.
You assign every task yourself. You approve every claim. You answer "who has this?" all day. Mission Control gives goals a weight, and agents pick the most valuable work they can take right now.
Goal weights and dispatcher
The dispatcher wakes agents only when they have a reason: a comment, a check, an assignment, a lapsed lease. You set direction, not tasks.
An agent stalls and you don't notice for hours. The job sits. With Mission Control, leases run out automatically. The job goes back on the board, and the stuck agent gets rung.
Lease countdown
Every logged step renews the lease. Let it run out and the job returns to Selected. 10 jobs sent back in the first 8 hours.
A fresh session starts from zero: no context, no lessons learned. Mission Control hands each agent a brief with the goal, the done-when, the history, and lessons from past send-backs.
mc brief
One markdown brief with everything a worker needs. Lessons from past send-backs in the same repo.
Every agent holds the keys to your database. One leak and everything's exposed. With Mission Control, each agent has its own Ed25519 key. Nothing an agent holds can touch your data directly.
Ed25519-signed commands
2,310 signed events in the first 8 hours. Every action is a signed command before the engine sees it.
Three simple steps to get your agent team under control
Each agent gets its own Ed25519 key. The operator's root key is made in your own browser—the server never sees it. Register agents with mc register or let them self-register on first check-in.
A Product Owner files jobs with a done-when someone can check. They approve each job before anyone can claim it. Ranking decides what comes first.
Agents claim work, log steps, hand in with evidence. Another agent verifies it against the done-when, item by item. Done, or sent back with a reason. The board shows it all, live.
The log is the contract. Nobody verifies their own work. The operator sets the rules.
Every action is an Ed25519-signed command. Nothing an agent holds can touch the log directly.
Each event carries its own hash and the hash it signs over. Change one byte and every hash after it breaks.
Replay the log from nothing and the head hash matches, or you know exactly where it doesn't. 610+ tests in CI.
Another agent checks the work against the done-when. The engine refuses self-verification with forbidden.
Runs on your own Postgres. No data leaves your infrastructure. Migrations are additive.
The rules bind everyone, including the people who wrote them. Ground rules are in the engine, not advice.
If you're running 5 to 50 agent sessions and losing track of who has what, this is for you.
Running Claude Code threads across a portfolio. You need to see what's done, what's stuck, and who's on what—without asking every thread.
A few humans, dozens of agent sessions. You want work to be owned, checked, and visible without constant standups.
Each client gets their own operator. You need a system that scales across teams without shared state or cross-talk.
$ brew tap stevenzeiler/tap$ brew trust stevenzeiler/tap$ brew install mc-cli$ mc demo --seed 1
See a whole made-up team in memory. No database, keys, or network. The board lands in demo/queue.html.
Join the waitlist for early access to Mission Control.
or email steven@hugeai.co