# Ovrin

> Secure cloud sandboxes for AI coding agents — one API call to provision an isolated Linux machine with the agent pre-installed, network locked down, and nothing left behind.

## What Ovrin does

Ovrin is an agent compute cloud. It gives AI coding agents (Claude Code, Codex, Gemini CLI, DeepSeek Harness) a real, isolated Linux environment to run in. Developers call one API to create a sandbox, run a prompt, collect the output, and destroy the environment — in seconds, not minutes.

Every sandbox is kernel-isolated with gVisor (syscall interception), namespace-scoped per tenant on k3s, and destroyed completely on kill. Egress is default-deny: only allowlisted hosts (e.g. api.anthropic.com, github.com, pypi.org) can be reached.

## Core capabilities

- **Compute sandboxes** — isolated Linux (Ubuntu), gVisor runtime, millisecond provisioning, disposable by design
- **Agent environments** — Claude Code, Codex (OpenAI), Gemini CLI, DeepSeek Harness — pre-installed, on the PATH, authenticated via env vars
- **Browser access** — Chrome and Playwright inside the sandbox boundary; agent can click, scrape, upload, download, take screenshots
- **Persistent memory** — `POST /memory`, `GET /memory/search` — context that outlives the sandbox; scoped per account
- **Payments** — each account uses its own encrypted Stripe restricted key through `StripeClient`
- **One SDK, one bill** — Python (`ovrin`) and TypeScript (`ovrin`) SDKs; single account for all capabilities

## API surface

```
POST   /sandboxes          create a sandbox
GET    /sandboxes/:id      inspect state
POST   /sandboxes/:id/run  execute a command (streaming stdout)
POST   /sandboxes/:id/files/write  write a file
POST   /sandboxes/:id/files/read   read a file
DELETE /sandboxes/:id      destroy (kill)

POST   /memory             store a memory
GET    /memory/search      semantic search over memories

POST   /secrets            write a credential (write-only vault)
GET    /keys               list scoped API keys
GET    /usage              metered compute and quotas
```

## Quick start (Python)

```python
import ovrin

client = ovrin.Client()

sandbox = client.sandboxes.create(
    template="claude-code",
    env={"ANTHROPIC_API_KEY": key},
    timeout=3600,
)

result = sandbox.run('claude "fix the failing auth test"')
print(result.stdout)
sandbox.kill()
```

## Security model

1. **gVisor** — userspace kernel intercepts all syscalls before the host sees them
2. **Tenant namespace** — per-account Kubernetes namespace with ResourceQuota and LimitRange
3. **Egress policy** — default-deny per sandbox; hosts allowlisted, not discovered
4. **API keys** — scoped, revocable; sandbox calls proxied with short-lived signed tokens

## Pricing

Self-service. No demo call, no sales gate, no waitlist. Metered by compute-second. See `/usage` for quota details.

## Links

- Homepage: https://ovrin.app/
- Agent environments: https://ovrin.app/agents
- Claude Code agent page: https://ovrin.app/agents/claude-code
- Other agent pages: https://ovrin.app/agents/codex · https://ovrin.app/agents/gemini-cli · https://ovrin.app/agents/deepseek-harness
- Docs: https://ovrin.app/docs — quickstart: https://ovrin.app/docs/quickstart — Python SDK: https://ovrin.app/docs/python-sdk — templates: https://ovrin.app/docs/templates
- Pricing: https://ovrin.app/pricing
- Security: https://ovrin.app/security
- Legal: https://ovrin.app/terms · https://ovrin.app/privacy · https://ovrin.app/dpa · https://ovrin.app/acceptable-use
- SDK (Python): `pip install ovrin`
- SDK (TypeScript): `npm install ovrin`