What is Squad? AI teammates on a dedicated OpenClaw computer, explained
Squad launched September 3, 2026: named AI teammates on one dedicated cloud computer per customer, on the ChatGPT or Claude plan you already pay for, $99 a month. What it runs on, what it does not publish, and how it differs from Grok Bot.
Squad (squad.so) launched on September 3, 2026 as “AI teammates to do your work”: named agents, each with one job, its own schedule, its own logins and its own model, running on one dedicated cloud computer per customer. The economic pitch is the reverse of Grok Bot’s: $99 a month to Squad, plus the AI plan you already pay for (ChatGPT at $100 to 200 a month is what they recommend), with “no per-task AI charges from us.” It grew out of MissionControlHQ, and its own resources page says squads “run on OpenClaw under the hood.” What it does not publish, anywhere: who hosts the computer, where, what survives an update, or any security page.
Where it came from
On January 31, 2026, its founder posted “This is my Mission Control: A Squad of 10 autonomous @openclaw agents. Led by Jarvis (my main @openclaw agent). They create work on their own. They claim tasks on their own,” and the post drew 782,000 views. MissionControlHQ productized the pattern through the summer, and on launch day the rebranded Squad named its foil directly: “Squad is exactly like Grok Bot but without any limitations of Grok Bot.” The founder also credited xAI with the timing: “When Grok Bot got launched, they kinda validated and educated the market about AI teammates for us.” Squad is, in other words, the open-harness answer to a closed product, sold as a finished service.
What a teammate is
The unit is a named agent with a single job. Squad’s example roster: one pulls overdue invoices from Stripe and drafts reminders; one triages the inbox and escalates; one handles newsletters and client correspondence; one screens applicants. An “AI Chief of Staff” interviews you, designs the roster and writes each job down as a procedure you can edit. Teammates are unlimited on the one plan, and can be hired, renamed or retired. Each gets its own model, so a research teammate can run on a different provider than the one answering mail.
The operating layer is what they call a mission control: a shared board every agent and human reads and writes, an @-mention that “triggers the mentioned agent’s run,” scheduled wakes (“agents wake every 30 to 60 minutes, check their lanes, act on what’s pending” plus named jobs on a timetable), inbound email as a trigger, approval gates (“anything that leaves the building waits for your yes”; outbound mail is draft-for-approval by default), a daily cap on unsupervised work, and a ledger of every run with its model, trigger and cost. A read-only dashboard can be shared with a client, which tells you who else buys this: the testimonials on the homepage are full of agencies and consultants onboarding their own customers.
The computer underneath
“Dedicated cloud computer per squad,” the homepage says: “real browsers, real files, real downloads. Nobody else’s work runs on it.” Within that one machine, teammates are separated by browser profile (“own logins and files, separate by default”) and by per-connection scopes on the 70-plus integrations (read-only, read and create, or full, held “no matter which agent calls”). The secret vault lives on that computer rather than in a shared database, and a teammate reads a secret at run time without posting it into a thread.
That is the same isolation unit as Grok Bot, one computer per account, refined with per-agent browser profiles and connection scopes. Squad’s own Grok Bot comparison leads with xAI’s “isolation is per user, not per Grok Bot” as the flaw and offers “per-agent boundaries” as the answer; the boundaries are profiles and scopes, not machines. Whether that is enough depends on what your teammates hold, and it is the question to ask before a teammate gets the payroll login.
What Squad does not say about the computer is as notable as what it does. There is no hosting provider named, no region, nothing on whether it is a VM or a container, nothing on what persists across their updates, no backup or snapshot story, no SLA (the terms say they “do not guarantee uninterrupted service”), and no security or status page. Cancelling puts “your workspace and data” on a deletion queue immediately; they tell you to export first. None of that is damning for a two-day-old product. It does mean there is nothing to compare an architecture against yet.
Bring your own AI plan
This is the spine of the offer. Supported plans include ChatGPT (recommended), Claude and Claude Code, Gemini, SuperGrok, Copilot, Kimi, MiniMax, Qwen and OpenRouter, and you can “stack multiple accounts with fallbacks.” Requests go “directly to the AI model providers you connect,” and Squad says it never trains on your data. Their own all-in figure is “$199 to 299 a month”: the $99 fee plus the plan. Two things the public pages leave unaddressed: how a consumer ChatGPT or Claude subscription is connected (the harnesses’ own sign-in flows are the obvious route), and whether running unattended agents on a consumer plan sits inside those providers’ terms. Both are worth asking before you build a business on it.
What it costs, and the small print
$99 a month flat, a seven-day money-back guarantee on the first payment, cancel from the dashboard. Add-ons: a real email inbox per teammate at $10 a month each, and research and scraping tools on credits (1,000 a month included, 5,000 for $5). Web, Telegram, Slack, Discord and email are the surfaces today; iOS is “soon.” Self-hosted or isolated deployments are “case by case.” The month-by-month comparison puts those numbers beside Grok Bot’s and beside running the same pattern yourself.
Where it sits
Three products pushed the “AI teammates” idea in four weeks: Grok Bot on August 11, OpenClaw 2.0’s own multi-agent and shared-session story on August 31, and Squad on September 3. Grok Bot sells a closed roster on its own models for a Cursor or SuperGrok subscription; Squad sells an office of employees on an open harness and your own plan, and hides the computer they run on; OpenClaw itself gives you the harness and most of the mission-control pieces to run the pattern on a computer of your own, board not included. Which of the three fits is mostly a question of whether you want to meet the computer.