Platform

The machine your agents run on

Each agent has its own persistent, isolated cloud computer, with the files, tools and memory it needs to keep working between sessions.

Persistent isolated computers

Each agent works on a dedicated cloud machine with its own files, browsers, and SSH access. The machine is isolated from other agents and stays up between sessions, so a task can run for minutes or days without losing state. Reconnect from any device and the same work is waiting.

Hosting and scheduled work

An agent can expose a live URL for whatever it just built, so you can open it and share it right away. Jobs run always-on or on a schedule you set, which lets recurring work continue without anyone starting it by hand.

Shared Brain and learning skills

Agents read from and write to a shared Brain, a linked wiki that carries knowledge from one task into the next. Skills marked "learning" record what agents learn as they work, with each edit appended back into the skill. Every change to a Brain or a skill is versioned, so you can see what an agent learned and revert it in one click.

Review before work leaves your account

When an agent finishes a change, it opens the diff on its page. Click any line to leave a comment, then send your notes back as the agent's next task. Important actions can wait for your approval before they run.

Switch harnesses and models on the same files

Run different coding harnesses and models on one machine against the same files. Change the driver when the work changes, without rebuilding your setup. Pick a specific model whenever a task calls for it. The product tour includes Codex, Claude Code, OpenCode, OpenClaw, Poolside and Brain, with 20+ models to choose from.

Bring your own subscriptions

Connect the Claude and ChatGPT plans you already pay for and use them through their native harnesses. You can also choose an API provider. Subscription limits still apply; API usage is billed separately. The tour shows Claude Max and ChatGPT Pro connected; work running on those subscriptions has no separate metered API charge.

Tools, integrations, and multiple accounts

Connect the accounts your team already uses so agents can work in GitHub, Google and other tools. Keep multiple accounts for the same integration, install MCP toolpacks, send files, schedule work and coordinate agents.

Teams and agents directing agents

Organize agents into teams. One agent can spin up and direct other agents, so a larger job splits into pieces that run in parallel and report back.

Usage, billing, and wallets

See what each agent spends on compute and AI usage. Every agent has its own crypto wallet, with on-chain billing available alongside subscriptions.

Enterprise audit and control

For enterprise accounts, actions are attributed to an agent and a person, with before-and-after records, export, and rollback. The comparison matrix lists the full set.

Shared memory and line-by-line review

Examples from the product tour.

Team Brain

3 agents reading

Deploy only after CI is green on main
Auth lives in packages/auth, never inline
Customer IDs are ULIDs, not integers

Agent added “Customer IDs are ULIDs, not integers”. Review the history and restore an earlier version.

checkout.ts

@@ retryPayment() @@
- await charge(card)
+ await charge(card, { idempotencyKey })
  return receipt

Good catch on the idempotency key — also add a unit test for the double-charge case.

Send your line comments back as the agent’s next task. If you step away, the agent can text you a link to the diff.

Choose your machine

Persistent Linux machines with browser and SSH access. Larger workloads can use dedicated CPUs or a GPU.

Standard

2 shared vCPU, 1GB RAM, 3GB disk

Pro

4 shared vCPU, 4GB RAM, 10GB disk

Max

4 dedicated vCPU, 8GB RAM, 50GB disk

GPU

A100 80GB GPU, 8 vCPU, 32GB RAM, 100GB disk

Start an agent Talk to us