# Agent as a Service — AI Agents in Sandy Sandboxes

Run hardened AI agents from hardened Linux sandboxes via a REST API.

## Base URL

https://agent-as-a-service.online/api

## Authentication

Authorization: Bearer YOUR_AGENT_API_KEY

Get API keys from `/account` and copy once at creation.

## Quick Start

1. Create a sandbox:

```bash
curl -X POST https://agent-as-a-service.online/api/sandboxes \
  -H "Authorization: Bearer YOUR_AGENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"persistent": false}'
```

Response returns `{ "id": "...", "status": "running" }`.

2. Launch an agent:

```bash
curl -X POST https://agent-as-a-service.online/api/sandboxes/<sandbox_id>/agent/run \
  -H "Authorization: Bearer YOUR_AGENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent": "claude-code", "prompt": "Build a simple REST endpoint", "model": "auto"}'
```

The endpoint streams SSE events with `event: status`, `event: output`, and `event: complete`.

3. Read a file:

```bash
curl -X GET "https://agent-as-a-service.online/api/sandboxes/<sandbox_id>/files/read?path=/workspace/out.txt" \
  -H "Authorization: Bearer YOUR_AGENT_API_KEY"
```

4. Write a file:

```bash
curl -X POST https://agent-as-a-service.online/api/sandboxes/<sandbox_id>/files/write \
  -H "Authorization: Bearer YOUR_AGENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"path":"/workspace/out.txt","content":"hello"}'
```

5. Terminate when finished:

```bash
curl -X POST https://agent-as-a-service.online/api/sandboxes/<sandbox_id>/terminate \
  -H "Authorization: Bearer YOUR_AGENT_API_KEY"
```

## Supported Endpoints

- GET `/health`
- POST `/sandboxes` / GET `/sandboxes` / GET `/sandboxes/{id}`
- POST `/sandboxes/{id}/exec`
- POST `/sandboxes/{id}/agent/run`
- POST `/sandboxes/{id}/agent/cancel`
- GET `/sandboxes/{id}/files` / read / write
- GET `/sandboxes/{id}/vnc`
- GET `/openapi.json` and `/openapi.yaml`

## Supported Agents

claude-code, codex, aider, openhands, factory-droid

## Notes

- Keep secrets in session-scoped server settings only; never expose raw model keys in public logs.
- Use concise prompts and cancel long-running tasks with `/agent/cancel`.