Cerver Compute · like OpenRouter, for computers

Let agents choose their computer.

Sandbox, GPU, desktop or your machine. One API, and the agent picks the right one at each step.

Get a computer

Paste the prompt into Claude Code, Codex or Cursor. Agents: cerver.ai/skill.md

One API across
E2BVercelDaytonaModalBoatCloudflaresoonAWSyour accountGoogle Cloudyour accountYour MacYour serversYour Linux box
One task"Find out why customers churned in October and put it in our board deck."
  1. 1Log in and export the accountsDesktop
  2. 2Score 2M support ticketsGPU
  3. 3Run the analysisSandbox
  4. 4Update the board deckYour laptop
Four computers. One API. The same calls on every one.
What runs where

Different work needs a different computer.

A video wants a big GPU. A chat reply wants almost nothing. A login wants a screen. Your files want your own Mac. Cerver sends each job to the computer that fits.

JobRuns onWhyTimeCost
Image generation50 product photos from one briefGPUA10 · 24 GBSeconds on a GPU, an hour on a CPU2 min$0.04
VideoA 30-second product videoLarge GPUA100 · 80 GBFrames need a lot of GPU memory8 min$0.33
Text and chat1,204 support repliesTiny sandbox1/16 vCPU · 256 MBThe model thinks; the computer waitsall day$0.18
Math and dataA 10-million-run simulationHigh CPU16 vCPU · 32 GBSplits across many cores90 s$0.02
Browsing and forms40 vendor forms behind a loginDesktop with a screen2 vCPU · 4 GBLogins need a real browser25 min$0.07
CodeFix a failing test, open a PRSandbox4 vCPU · 8 GBIsolated, safe to run anything6 min$0.03
Private filesUpdate the board deckYour machineYour Mac · your planFiles never leave your Mac3 min$0
Research at scaleRead 1,000 pages50 in parallel50 small sandboxesMinutes instead of hours4 min$0.02

Example jobs. Costs are time × the provider's published hourly rate (Nov 2026).

Same computer for every job vs. the right one for each

JobMachine-hours / moSame computerRight computerCost on it
Total a month
Same computer
Right computer for each step
less a month, saved on your provider bill.

Prices from the providers' pricing pages, November 2026, per hour with the CPU busy: Modal A100 80 GB $2.50, Modal A10 $1.10, Modal 16 vCPU $0.63, E2B 4 vCPU $0.33, E2B 2 vCPU $0.17, Cloudflare lite $0.007, your Mac $0. Cloudflare and Vercel bill CPU only while active, so idle time costs less. Example workload in machine-hours; 50 sandboxes for 4 minutes count as 200 machine-minutes.

Compute pools · soon

At scale, keep computers warm. A router sends every task to the best fit.

Keep a few computers of each type running. The router reads each task, picks the pool that fits, and decides what to do when it is full: wait, burst to on-demand, or take the next best pool. Use plain rules, or let an agent route.

Incoming

0 waiting

Router

Pools

DesktopE2B · screen + browser
GPUGPU box · Modal
FastVercel · Cloudflare
Privateyour Mac minis
warm, idlebusyon-demand burstcold starts avoided: 0
# one policy for the whole fleet
pools:
  desktop:  { on: e2b,          warm: 3, max: 6 }
  gpu:      { on: gpu-box,      warm: 1, max: 2 }
  fast:     { on: vercel,       warm: 4, max: 8 }
  private:  { on: my-mac-minis, fixed: 4 }
route: agent     # or: rules (by tagged need)
when_full: cheapest of [wait, burst, next pool]
  • Rules or an agent. Rules match a tagged need. The router agent reads the task itself and explains every choice.
  • Your own machines take the steady load. Cloud computers only cover the spikes.
  • See every pool. Busy, idle, waiting and cost, so you size the pools to your traffic.
01

The right computer for each step

Desktop, GPU, sandbox or your own machine. Start each step on the one that fits, and move a session between the computers on your account without losing a step. GPUs through Modal are coming soon.

02

Watch and take over

See each computer's screen live. When the agent is stuck on a login or a 2FA code, click in, type it, and hand it back.

03

One view, every computer

Each step priced on the computer it used, rolled up per agent and per customer. A $20 cap stops at $20.