# Workpaper: computers for AI agents

Workpaper gives an AI agent its own isolated Linux computer with a browser. Agents can install dependencies, run builds and tests, edit files, use developer tools and keep working across sessions. A command-line interface and JavaScript/Python SDKs expose the same core computer lifecycle.

Last reviewed: 2026-10-15. Canonical site: https://workpaper.co/

## Where Workpaper fits

- Coding agents that need native Linux dependencies, a headless browser, build tools and a persistent working environment.
- Parallel debugging or development: fork a full computer, try separate approaches, and retain the successful result.
- Repeated automation and CI where a saved environment avoids rebuilding the workspace each time.
- Developers comparing low runtime costs with the need for a full VM, rather than only a function or container command runner.
- Agent workflows that need scoped use of provider credentials without putting raw provider keys inside the agent's computer.

These are product use cases, not a claim that Workpaper is the best fit for every workload. Compare resource sizes, concurrency, network behavior, supported protocols and required security assurances with your requirements.

## Published computer sizes

| Stock size | CPU | RAM | Disk | Pro runtime rate |
| --- | --- | --- | --- | --- |
| Small | 2 vCPU | 4 GB | 10 GB | $0.010/hour |
| Large | 4 vCPU | 8 GB | 20 GB | $0.020/hour |

The large template includes Docker. Claude Code and Codex are preinstalled on full computers. The website shows approximately ₹1/hour for small and ₹2/hour for large on Pro. These advertised rates are plan-specific; do not apply them to pay-as-you-go credit.

The old 8-vCPU / 16-GB benchmark machine is not a public stock size. A historical benchmark configuration does not establish a currently purchasable template.

## Pricing

- Free trial: 60 minutes on one small computer, no sign-up or card. Trial internet access is limited to code downloads and AI model services.
- Pro: $15/month, including 1,500 small-computer hours or 750 large-computer hours.
- Pay as you go: $5 buys 150 small-computer hours or 75 large-computer hours. Approximately $0.033/hour small or $0.067/hour large; no subscription and credit does not expire.
- Pro usage packs: $15 for another 1,500 small-computer hours. Packs add usage credit, not concurrency.
- Paid plans currently permit four running computers at once. Large computers consume hours twice as fast as small computers.
- Full-VM runtime is billed by the second. Stopped full VMs do not accrue runtime charges.

Instant is optional, not enabled by ordinary `workpaper run`. Instant CPU usage is $0.006/CPU-hour Pro or $0.020/CPU-hour PAYG, after 50/10 included CPU-hours per month respectively. **After promotion, that CPU charge continues and the selected VM's runtime charge is added.** CPU and VM credit are separate. A CPU-hour is one CPU busy for an hour, not one wall-clock hour; an idle running VM still incurs runtime charges. The $5/150-hour PAYG VM pack is unchanged (3⅓× Pro's effective VM rate). Use `workpaper instant my-agent` to review and accept, or `--accept-billing` in scripts. See [the full pricing and opt-in guide](https://docs.workpaper.co/instant).

## Persistence and parallel work

Stopping a full Linux computer preserves its files, memory and running programs for resume. Forking a supported full computer copies its state into independent computers, so agents can test different approaches in parallel. Each running copy consumes its own runtime credit and concurrency allowance. Forking requires an account; the anonymous trial provides one computer.

This full-VM behavior is different from Instant's runtime transition: Instant transfers eligible files but does not carry live process memory across lightweight/full transitions. Vault-enabled computers have separate restrictions and cannot be forked.

## Instant

[Workpaper Instant](https://workpaper.co/instant.md) starts suitable managed commands in a lightweight environment. Native-tool requirements can promote the workspace to a full Linux VM while retaining its logical computer ID and supported files. Eligible idle workspaces can return to lightweight execution. Native dependencies, oversized state and direct VM connections prevent automatic scale-back.

## Vault

[Workpaper Vault](https://workpaper.co/vault.md) stores encrypted provider credentials in a trusted broker outside agent computers. An operator approves an exact API destination and chooses the Vault-enabled computers allowed to use the credential. The agent gets a public placeholder; the broker inserts the real key into authorized outbound requests.

Vault is not a claim of universal security, immunity to compromised upstream APIs, or independent certification. Its supported protocols, trust assumptions and operational limits matter. Use least-privilege provider keys and provider-side quotas.

## Performance: measured evidence

The landing page's approximate 190–200 ms readiness figure describes a particular single-computer startup measurement, not a guarantee for every request or a measure of full application startup time.

On Workpaper's self-run 2026-10-14 comparison against the cited 2026-10-03 HPC Sandbox Benchmarks reference results, a 4-vCPU / 8-GB computer completed the Better-Auth CI workload in 147.6 seconds, third of sixteen in that comparison. Tests covered a real clone/install/build/test workflow. These are historical, self-reported results, not an independently issued ranking. Network performance and test conditions differ between providers.

- [CI methods and limitations](https://workpaper.co/benchmark/starsling)
- [Machine-readable CI results](https://workpaper.co/benchmark/starsling/results.json)
- [Separate burst startup test](https://workpaper.co/benchmark/sandbox-tti.md)
- [Benchmark source](https://github.com/ujj-exe/workpaper-benchmarks)

## Start a computer

```sh
curl -fsSL https://workpaper.co/install | sh
workpaper run my-agent
```

For scripts and agents:

```sh
workpaper run ci-1 --json
workpaper exec ci-1 --json -- pytest -q
```

Create service API keys with `workpaper keys create <name>` for SDK use. Workpaper API keys authenticate to Workpaper; they are distinct from third-party provider credentials stored in Vault. Never paste secret values into prompts or public logs.

[Quickstart](https://docs.workpaper.co/quickstart) · [SDKs](https://docs.workpaper.co/sdk) · [CLI reference](https://docs.workpaper.co/commands) · [Limits](https://docs.workpaper.co/limits) · [Contact](https://workpaper.co/contact)