OpenClaw mission control: running a squad of agents on a computer of your own

What the phrase means

The term was made famous by a January 2026 post, “This is my Mission Control: A Squad of 10 autonomous @openclaw agents … They create work on their own. They claim tasks on their own,” which drew 782,000 views and, eight months later, became a product. That company’s own definition is a fair one: “the operations layer that turns several AI agents into a team,” with five tests. State survives weeks, not sessions. Agents wake agents, by mention, schedule or event. Humans hold the gates on anything consequential. Cost is attributed per run and per agent. And the view is shareable without a terminal. Hold OpenClaw up against those five and you get a precise picture of what is native and what you build.

What OpenClaw 2 gives you out of the box

Many agents, one Gateway. Each agent has its own workspace (its own SOUL.md, AGENTS.md, notes), its own state directory with its own SQLite session store and auth profiles, and its own tool allow and deny lists and sandbox settings. Chat accounts are routed to agents by bindings, so one Telegram account can talk to one agent and a Slack workspace to another, or one WhatsApp number can split its senders between agents. How multi-agent works covers the mechanics.

Agents that use each other. The session tools let an agent run another session on the same Gateway and wait for the answer (sessions_send), start a background child for a task and get a run id back (sessions_spawn), and end its own turn so the child’s completion can be announced (sessions_yield). Cross-agent access is on by default and can be restricted pair by pair. One precise limit from the docs: agents cannot autonomously wake each other; a parent that spawned or is watching a session gets a state-change notice, and acts on it on its next turn. That is the difference between “your researcher messages your coder” and a board where a mention starts a run.

Schedules and standing work. Automations cover one-shot reminders, recurring intervals, cron expressions and webhook triggers, with output delivered to a chat or a URL; the heartbeat is a built-in “check your lane” turn every 30 minutes; standing orders in AGENTS.md give an agent permanent operating authority for a defined program; Task Flow keeps “a durable record of multi-step work with its own status, JSON state, revision counter” across restarts, inspectable with openclaw tasks flow list.

Gates. Permission modes (read-only, guarded, workspace, full) decide whether a command runs, asks a human, or asks a model reviewer with a human fallback; exec approvals allowlist the rest; and since 2.0 an agent’s changes to its own configuration are proposals only the owner can approve. For a team of people rather than agents, 2.0 also added operator roles and shared sessions with an immutable creator and an assignable owner, the “who did what” that a shared Gateway needs.

What you build, or buy

Against the five tests, OpenClaw passes “state survives weeks” (everything is files and SQLite on your disk), “humans hold the gates” and, with automations, “agents wake on schedule.” The two it does not pass alone are the board and the ledger-as-a-view. There is no shared task list that every agent reads and writes and that starts a run when one agent mentions another; there is no single screen of runs with model, trigger and cost per row. Usage is recorded per session in the store and openclaw tasks audit lists background work, so the data exists; the view does not.

The do-it-yourself version is older than the products. The board is a folder or a git repository the agents share: a Markdown file per task, status in the front matter, claimed by whichever agent picks it up, with commit history as the audit trail. The wakes are automations: each agent scheduled to read the board every so often, the heartbeat as the fallback. Handoffs are sessions_send with a task id. Gates are permission modes plus a rule in each agent’s standing orders that outbound mail and payments are drafts until approved. The view is whatever reads the folder: a static page, a spreadsheet, the repository’s own web UI. The January post that started all this was running on roughly that shape. Open-source dashboards that wrap it appeared within weeks of it.

What the hosted versions charge for

Squad is $99 a month plus the AI plan you bring (they recommend ChatGPT at $100 to 200), for the board, the @-mention wakes, approval gates, the cost ledger, a shareable client dashboard and 70-plus integrations with per-agent scopes, on a dedicated computer per squad whose provider and persistence they do not publish. Grok Bot comes bundled with Cursor and SuperGrok plans from Cursor Pro at $20 a month upward, with weekly usage and then metered billing, on one shared computer per account and xAI’s models only. The month-by-month comparison prices the three routes side by side. The summary: the hosted products sell the board and the polish, and the open route sells you the computer and the harness and leaves the board to you.

A build order that works

1. One agent, one job, until it is reliable. Most “squads” that fail did this step last.
2. Add a second agent with a distinct job and its own workspace (openclaw agents add), and give it a different chat binding so you can tell who is talking.
3. Make the shared folder the board and put its conventions in both agents’ AGENTS.md as standing orders.
4. Schedule the wakes with automations, staggered so they do not all fire on the hour; let the heartbeat be the safety net, not the scheduler.
5. Set permission modes so nothing leaves the building without you, and keep it that way until the ledger shows a month of boring runs.

Where the computer comes from

Every version of this needs a machine that is always on and belongs to the squad. Run it yourself on a VPS, or use managed OpenClaw hosting and skip the administration. On an Everpod pod OpenClaw runs as the project ships it, so its multi-agent features, automations and approvals are the owner’s to use as on any other machine; the board is yours to choose, and we do not add one. A pod is one customer’s computer, $29 a month with up to $10 of model usage included, which for a squad that runs on your own ChatGPT or Claude plan is mostly the computer. The thing to remember, whichever route you take: the products above have pricing pages because a working multi-agent setup is more than the sum of its tools. Budget a real weekend.