■ Full VMs for agents · early access

Real computers for your agents.

Full Linux virtual machines, not containers. Billed by the second, nothing to start one, and when we're full you wait in line with an ETA instead of getting an error.

POST /v1/sandboxes

$ curl https://raw.computer/v1/sandboxes \
  -H "Authorization: Bearer $RAW_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"class":"std","size":"4x8"}'

Agents: start at raw.computer/llms.txt

Machine: Full Linux VM
Billing: Per second
Start fee: $0
When full: Queue with an ETA

Built for software that works on its own.

01
Full Linux VMs

Each sandbox has its own kernel: systemd, Docker, nested virtualization, a persistent disk, snapshots and forks. If it runs on a Linux server, it runs here.
- systemd
- dockerd
- /dev/kvm
- snapshot
- fork
02
No start fees. A queue, not an error.

Billed per second while it runs; creating, resuming and forking cost nothing. When we're full, your start waits in line and tells you its position and ETA.
- start_fee: $0
- billing: per_second
- 202 queued
- queue.eta
03
Built for agents

An HTTP API with typed ids and idempotency keys; a CLI whose every command answers in JSON and also runs as an MCP server; a skill for Claude Code, Codex and Cursor; llms.txt and every docs page as Markdown. Every error names the next step.
- /v1
- raw mcp
- SKILL.md
- llms.txt
- next_step

Create, run, snapshot, fork.

Four calls take an agent from nothing to two machines in the same state. Step through them to see each request and its answer.

This is a demo: it calls nothing, and the answers are written by hand in the API's shapes.

raw.computer/v1 Demo

A full VM with its own kernel, systemd and Docker. With ?wait=ready the answer comes when it runs.

$ curl -X POST "https://raw.computer/v1/sandboxes?wait=ready" \
  -H "Authorization: Bearer $RAW_API_KEY" \
  -H "Idempotency-Key: 0f9d6c1e-52b4-4a8e-9c3f-6e2d1b7a8c90" \
  -d '{"class":"std","size":"4x8"}'

HTTP/1.1 201 Created

{
  "object": "sandbox",
  "id": "sbx_ny1_01m6xxh3k2q8v7c4t9r5wdz6ya",
  "name": null,
  "class": "std",
  "size": "4x8",
  "state": "running",
  "created_at": "2026-11-08T17:02:11.000Z"
}

Priced per second. Compared in the open.

Two classes, one linear price list, no plans and no start fees. The calculator prices your workload here and at other providers, from their published list prices.

std · 4 vCPU / 8 GiB: $0.032/h; agents and dev boxes
perf · 4 vCPU / 8 GiB: $0.048/h; CI, dedicated cores

Full price list

Monthly cost calculator

Your workload, priced at our list prices and at other providers'.

Sessions per month

Minutes per session

Size (vCPU / GiB)

Providers without this size are listed last.

CPU busy 25% Only Vercel bills CPU by busy time; everyone else, us included, bills the machine while it runs.

Monthly cost of the workload by provider, cheapest first
Provider	Per month	vs std
Agent37 (VM) disk billed extra	$30.14	0.9×
raw.computer std shared vCPU, per second	$32.00	1.0×
Lizard $19.00/month plan, counted as usage	$32.40	1.0×
boat.dev $20.00/month plan, counted as usage	$36.00	1.1×
raw.computer perf dedicated cores, per second	$48.00	1.5×
Vercel Sandbox $20.00/month plan, counted as usageCPU billed when busy	$297.60	9.3×
E2B	$331.20	10×
Daytona disk billed extra	$331.20	10×
Modal after $30.00 of free monthly credit	$445.92	14×

Each provider's name links to its own pricing page, read on 2026-11-08. Starts are assumed spread evenly over 30 days; where starts are charged, bursts cost more. Taxes excluded.

Benchmarks

Measured, not claimed.

You won't find speed numbers here yet. We publish a number only after our own harness has measured it on the machines we sell, together with the method and the raw results.

How we will measure

Real computers for your agents.

Built for software that works on its own.

Full Linux VMs

No start fees. A queue, not an error.

Built for agents

Create, run, snapshot, fork.

Priced per second. Compared in the open.

Monthly cost calculator

Measured, not claimed.

Give your agent a computer.