# cpuos

> cpuos.si gives every AI agent its own Firecracker microVM in under a second: run the code it wrote, drive a headless browser, test a repository. Hard-isolated, billed per second, free while paused, hosted in the EU or on your own servers.

- Positioning: The CPU OS for super intelligence. Models think on GPUs. Agents act on CPUs. SI stands for super intelligence; .si is also Slovenia's country domain. cpuos is where agents act, gpuos is where models think.
- Status: early access. Teams join the waitlist at https://cpuos.si/#access.
- What it is: an API that gives every AI agent its own Firecracker microVM sandbox, restored from a warm snapshot in under a second, to run the code it wrote, drive a headless browser or test a repository.
- Isolation: one Firecracker microVM with its own Linux kernel per sandbox, started through jailer with seccomp and cgroups, on dedicated bare-metal hosts. No container shares a kernel with another customer.
- Network: egress policy per sandbox (internet on, allowlist, or off). Private ranges, SMTP and the cloud metadata address are always blocked.
- Lifecycle: create, exec (streamed stdout and stderr, exit code, timeouts), files (read and write), ports at private HTTPS URLs (https://<port>-<id>.sbx.cpuos.si), pause and resume (memory and disk snapshot, no CPU or RAM billing while paused).
- Templates: Python (CPython 3.13 · uv · numpy · pandas · matplotlib); Node (Node 24 · pnpm · Bun · TypeScript); Browser (Chromium · Playwright · CDP · live view); Dev box (git · Python · Node · Go · Rust · build-essential); Custom (Your Dockerfile or OCI image).
- Interfaces: TypeScript SDK (`@cpuos/sdk`), Python SDK (`cpuos`), REST API, MCP server, adapters for the OpenAI Agents SDK, Vercel AI SDK and LangGraph.
- Hosting: hosts in Germany and Finland, an EU company, DPA included. Or self-hosted.

## Guides

- [What is an AI agent sandbox?](https://cpuos.si/guides/what-is-an-ai-agent-sandbox): An AI agent sandbox is an isolated, disposable computer where an agent runs the code it writes. What it must block, what agents need from it, how to pick one.
- [How to run LLM-generated code safely](https://cpuos.si/guides/run-llm-generated-code-safely): Secure code execution for LLMs: the threat model, why exec() and plain containers fall short, and how gVisor and Firecracker microVMs compare.
- [Firecracker vs Docker for AI agents](https://cpuos.si/guides/firecracker-vs-docker-for-ai-agents): Firecracker microVMs vs Docker containers for running AI agent code: kernel isolation, startup time, snapshots, density, operations and when each makes sense.
- [Build a code interpreter with open models](https://cpuos.si/guides/build-a-code-interpreter-with-open-models): Build a ChatGPT-style code interpreter with an open model on gpuOS and a cpuos sandbox: tool schema, agent loop, files, charts and limits, in Python.
- [An EU-hosted E2B alternative for agent sandboxes](https://cpuos.si/guides/e2b-alternative-eu): Looking for an E2B alternative hosted in the EU? How to compare agent sandbox providers on GDPR, data residency, isolation, pricing and self-hosting.
- [A browser automation sandbox for AI agents](https://cpuos.si/guides/browser-automation-sandbox-for-agents): Why browser agents need isolation, and how to run headless Chromium and Playwright for an AI agent inside a disposable microVM sandbox.

## Integrations

- [OpenAI Agents SDK](https://cpuos.si/integrations/openai-agents-sdk): Give an OpenAI Agents SDK agent a sandboxed code tool: a function tool that runs code in a cpuos microVM, with any OpenAI-compatible model.
- [LangChain and LangGraph](https://cpuos.si/integrations/langgraph): Add a sandboxed code execution tool to LangChain and LangGraph agents, with the model on any OpenAI-compatible endpoint.
- [Vercel AI SDK](https://cpuos.si/integrations/vercel-ai-sdk): Run AI SDK tool calls in a Firecracker sandbox from Next.js or Node, with the model on any OpenAI-compatible provider.
- [LlamaIndex](https://cpuos.si/integrations/llamaindex): Give a LlamaIndex FunctionAgent a code execution tool that runs in a cpuos microVM, next to your RAG tools.
- [CrewAI](https://cpuos.si/integrations/crewai): Give CrewAI agents a custom tool that runs code in a Firecracker microVM instead of a local Docker container.
- [MCP (Model Context Protocol)](https://cpuos.si/integrations/mcp): Expose cpuos sandboxes as MCP tools so Claude, Cursor and any MCP client can run commands and files in an isolated microVM.
- [Playwright](https://cpuos.si/integrations/playwright): Run Playwright and headless Chromium for browser agents inside a disposable microVM, using the cpuos Browser template.
- [Jupyter](https://cpuos.si/integrations/jupyter): Run JupyterLab in a cpuos sandbox: a live notebook the user can open while the agent analyzes data, isolated in a microVM.

## Pricing

- Usage, billed per second while a sandbox runs: $0.045 per vCPU-hour, $0.005 per GB-hour of RAM, $0.10 per GB-month of snapshot storage. A 2 vCPU / 4 GB sandbox costs about $0.11 per running hour. Paused sandboxes only pay snapshot storage.
- Free: $0 per month. $5 of usage every month; 5 concurrent sandboxes; Up to 2 vCPU / 4 GB each; 1-hour sessions, pause and resume; Python, Node and Browser templates.
- Team: $49 per month or $490 per year. $49 of usage included; 50 concurrent sandboxes; Up to 8 vCPU / 16 GB each; 24-hour sessions, 50 GB of snapshots; Custom templates from a Dockerfile; Egress allowlists.
- Business: $249 per month or $2,490 per year. $249 of usage included; 500 concurrent sandboxes; Up to 16 vCPU / 32 GB each; 7-day sessions; SSO (SAML / OIDC) and audit export; Dedicated hosts, 99.9 % SLA.
- Self-hosted: the same software on your own servers, per-host license.
- Prices in USD, excluding VAT.

## Sibling product: gpuos (models)

cpuos runs agent actions on CPUs. For the model itself, the sibling product gpuos turns your own NVIDIA GPUs into a private super intelligence cloud: open models behind one OpenAI-compatible API at https://gpuos.si/v1. The two products have separate accounts and work with any provider; together, the model thinks on gpuos and the agent acts in cpuos.

- [gpuos](https://gpuos.si): private SI cloud on your own GPUs.
- [gpuos llms.txt](https://gpuos.si/llms.txt): the gpuos reference for LLMs.
- [Self-host an OpenAI-compatible API](https://gpuos.si/guides/self-host-openai-compatible-api)
- [gpuos model catalog](https://gpuos.si/models)

## Optional

- [Full text for LLMs](https://cpuos.si/llms-full.txt)
- [Home page](https://cpuos.si)