Two products.

Compute gives an agent every computer it needs. Agents runs, schedules and caps the work on them. Use one or both.

Cerver ComputeComputers for your agents, on your own provider keys. From $20 a month. Cerver AgentsRun, schedule and pay for agents. From $29 a month.

Cerver Compute

Every computer, one API.

Sandboxes, GPUs, desktops and your own machines behind the same calls. The agent picks, and can change its mind mid-task.

Sandboxes

Name a provider and get a computer. The calls to run code, write files and stream output are the same on every one.

E2B and Vercel today
Your own Mac, Linux box or server with one install command
Cloudflare, Modal GPUs and Daytona workspaces are in testing

Computers in the docs

# same call, any provider
POST /v2/sessions
{ "compute": { "provider": "e2b" } }

POST /v2/sessions/:id/run
{ "code": "npm test" }
→ exit_code 0 · 41 passed

Switch mid-task

Move a running session to another computer on your account. The transcript and every step so far come along, so the agent carries on where it was.

Start in the cloud, finish on your laptop where the files are
Laptop closing? Move the work to a machine that stays on
Or give an ordered list and Cerver takes the first one that is up
Moving straight to a fresh cloud computer by provider name soon

Switching in the docs

POST /v2/sessions/:id/change-compute
{ "compute": { "compute_id": "comp_macmini" } }

# or first available, in order
{ "compute_priority": ["comp_macmini", "comp_laptop"] }

Live screen

Watch the agent's computer while it works. When it hits a login or a 2FA code, click in, type it, and hand the machine back.

One element in your page, two lines
Viewer tokens are per session, short-lived and view-only by default
Video goes straight from the machine to the browser

Live screen in the docs

<!-- your page -->
<script src="https://cerver.ai/embed/cerver-screen.js"></script>
<cerver-screen session="SESSION_ID" token="cvs_…"></cerver-screen>

# your server mints the token
POST /v2/sessions/:id/screen/token
{ "control": true, "ttl_seconds": 900 }

Your compute pool routing soon

Every computer your agents can use, in one pool: the provider accounts you connected and your own machines. See what is on, what it costs and who bills you. Soon, keep a few of each kind warm and let a router send every task to the best fit.

See how routing works

incoming            → routed to
render 40 thumbnails  a GPU         2 warm
reply to a chat       a small sandbox 6 warm
log in and export     a desktop     1 warm
read ~/reports        your Mac      online

Cerver Agents

Run agents on demand.

Every run is kept, the gateway lets any agent or model drive it, it can run on a schedule, and it stops at the cap you set.

Sessions

A session is one piece of work: the transcript, the tool calls, the computer it ran on and what it cost. Nothing is lost when the machine goes away.

Read any run back, search across all of them
Resume a session later on a different computer
Share a read-only link

Sessions in the docs

GET /v2/sessions/:id
{
  "status": "idle",
  "provider": "e2b",
  "transcript": [ … 42 entries … ],
  "metrics": { "cost_estimate_usd": 0.31 }
}

Gateway

One door for every agent and every model. Pick the agent and the model per session, change either mid-task, and see what each developer and each customer spent.

Claude Code, Codex and other CLIs, or a plain model call
Change the model without touching your app, or set a routing policy
Route a team's Claude Code and Codex through it with one command
Metered per developer, capped, secrets redacted
Runs on your Claude or ChatGPT plan, or on an API key from your vault

Agents and models in the docs

# pick the agent, or just the model, mid-session
POST /v2/sessions/:id/switch-tool
{ "cli_tool": "codex" }
{ "cli_model": "claude-haiku-4-5-20251001" }

# route a team's Claude Code and Codex through Cerver
cerver connect --project acme
cerver gateway on    # metered, capped, redacted
cerver gateway off   # back to your subscription

Crons & workflows

Agents that run on their own. Give a schedule and a prompt, or a workflow of fixed steps. Every firing is a session you can read.

Standard five-field schedules, in UTC
Run an agent, a workflow, or call a webhook
Pick the computer per cron

Crons in the docs

POST /v2/projects/:slug/crons
{
  "name": "nightly-report",
  "schedule": "0 6 * * *",
  "prompt": "Summarise yesterday's signups.",
  "harness": "claude"
}

Chat in your product

A chat for your users in two lines. Cerver stores the conversations, streams the replies and tells you what each customer costs.

Chat in the docs

<script src="https://cerver.ai/embed/cerver-chat.js"></script>
<cerver-chat pk="pk_live_…"
  system-prompt="You answer questions about our docs.">
</cerver-chat>

Cost & caps

Every step is priced on the computer and model it used, and rolled up per agent and per customer. A cap is a hard stop, not an alert.

Account limit, and a limit per agent type
Usage per customer, from one call

Cost and caps in the docs

PUT /v2/account/spending-limit
{ "limit_usd": 20 }

GET /v2/account/usage-by-customer
→ spend per external_user_id