Detach intent from execution

Your agent says what it needs. Cerver finds compute, runs it, keeps the session. If your machine is busy, Cerver routes to the next one. If it's 3am, the session waits. Intent and execution are decoupled — your agent never thinks about infrastructure.

Read docs See the model

Usage

your_app.ts
const session = await fetch("/gateway/sessions", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    task: "Boot a preview and run tests",
    requirements: { runtime: "node", timeout_minutes: 20 },
    policy: { mode: "balanced" }
  })
})

await fetch(`/gateway/sessions/${sessionId}/run`, {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    code: "npm test",
    timeout: 600
  })
})

session result
{
  "session_id": "sess_123",
  "provider": "vercel",
  "compute_id": "cmp_123",
  "stdout": "test suite passed",
  "metrics": {
    "provision_time_ms": 812,
    "last_exec_latency_ms": 1432,
    "cost_estimate_usd": 0.01
  }
}

What Cerver does

Automatic failover

If your Mac is busy running 3 agents, Cerver routes the next task to your Mac Mini. If that's busy too, it goes to Vercel. You set the priority — Cerver handles the routing.

Sessions that survive

Pause a session, close your laptop, come back tomorrow. The transcript, state, and context are preserved. Resume on the same machine or a different one.

Agents spawn compute

An agent running a task can request more compute mid-work. Spawn a child session for parallel tests. The parent-child relationship is tracked.

Your machines, your cloud

Register your laptop, your server, or your team's machines. Add Vercel or E2B with your own credentials. Cerver routes across all of them.

One API for everything

Create session. Run code. Pause. Resume. Spawn. That's it. Your agent talks to one endpoint — never to providers directly.

Built for AI agents

API reference and docs designed for AI agents to read. An agent can learn the API, create sessions, and manage compute without human help.

Startup profiles

Warm local compute fast

Warm remote compute balanced

Fresh remote compute cold start

Different computers wake up differently. Cerver keeps the session API stable and reports the startup profile back as metrics.

How it works

Register compute

Add cloud providers or connect your own machines via Cerver Connect.

Open a session

Describe what you need — runtime, persistence, timeout. Cerver checks your compute priority and picks the first available machine.

Run, pause, resume

Run code. When done, pause — compute is freed, session is saved. Resume later on the same machine or a different one. Context follows the session, not the machine.

Scale by spawning

Need more compute? Spawn child sessions. They inherit the parent's context and run in parallel. Results flow back when done.