# bead

> Sandboxes and cloud computers for AI agents: Linux VMs, fleets, swarms and agent runners, paid from prepaid credit.

- AI agents: read https://bead.run/SKILL.md.
- [Agent onboarding and authentication](https://api.bead.run/auth.md): free agent signup, GitHub consent and bearer keys.
- API: https://api.bead.run (CLI default; override BEAD_API_URL or --api-url).
- MCP: https://api.bead.run/mcp; Streamable HTTP.
- Dashboard: https://app.bead.run (humans connect a Claude or ChatGPT subscription under Agents & credentials).
- Install: https://bead.run/install
- Discovery: https://bead.run/llms.txt
- Skill: https://bead.run/SKILL.md; MCP extension `io.modelcontextprotocol/skills`, URI `skill://bead/SKILL.md`; prompt `bead`.
- OAuth discovery: https://api.bead.run/.well-known/oauth-protected-resource/mcp.
- CLI: Rust binary `bead`; `bead --help`, `bead --json` for automation.

Create through POST /v2/sandboxes, run through envd or runner tools, then delete
through DELETE /sandboxes/{id}. Credit is a hard cap; export before expiry.

Pricing: GET /v1/pricing returns the current USD rates and machine shapes. POST /v1/pricing/estimate prices a workload before you create it; `cpu_busy_share_bps` is the busy CPU share from 0 to 10000. Humans add prepaid credit through POST /billing/purchases, POST /billing/purchases/{id}/checkout and GET /billing/purchases/{id}; card and supported USDC payments fund the same credit balance. Keep purchase secrets out of URLs.

Agents sign up free, no card: `bead signup`. $10 credit to start.
The CLI solves a proof of work and saves its org key privately; MCP `account_create` and `POST /v1/signup` use the same path.
People sign in and add credit, or claim their agent’s org for dashboard access.

Interactive MCP consent uses GitHub sign-in at https://app.bead.run.

Install the CLI:

```sh
curl -fsSL https://bead.run/install.sh | bash
```

SDK quick start:

```sh
npm i @beadrun/sdk
pip install beadrun
```

```python
from beadrun import Sandbox
```

```typescript
import { Sandbox } from "@beadrun/sdk";
```

Machine sizes: xs=1vCPU/2GiB/20GB, small=2/4/40, medium(default)=4/8/40,
large=8/16/80, xlarge=16/32/160. Create body accepts size; CLI --size; SDK
Sandbox.create(size="medium") / Sandbox.create({size:"medium"}); MCP
bead:sandbox_create body.size. Fleet request.size and swarm size; fork inherits.
Trial supports xs/small. E2B cpuCount/memoryMB round up to the smallest fitting
size (400 above xlarge or conflicting with size). Memory snapshots require the
source size. Only small/medium stay warm. AX41 refuses xlarge (<16 host threads).
GET /v1/pricing is the size/rate/concurrency catalogue. Named estimates:
POST /v1/pricing/estimate {"size":"medium","duration_seconds":3600,"storage_bytes":0}.
Disk is a capacity limit, billed only on unique bytes actually retained.

<!-- generated-pricing:start -->
Prices are USD, before tax. CPU $0.0045/busy vCPU-hour
and configured memory $0.00225/GiB-hour while running.
`GET /v1/pricing` supplies the current per-size prices.
Unique retained storage is $0.06/GiB-month (730 hours).
Templates with a live desktop, including computer and devbox-computer, add
$0.04/running hour. Paused and evicted machines pay storage only.
CPU uses measured busy cgroup time; idle CPU is free. Memory uses configured
capacity regardless of resident pages. Storage uses unique retained bytes, including
snapshots and goldens; disk capacity is a limit. Creates, forks, snapshots,
pause/resume and EU egress have no operation charge. Existing prepaid credit
remains USD credit; prices apply from activation onward, with no retroactive re-rating.

| Size | vCPU | Memory (GiB) | Usable disk (GiB) | Busy / waiting ($/h) |
|---|---:|---:|---:|---:|
| xs | 1 | 2 | 20 | 0.00900 / 0.00450 |
| small | 2 | 4 | 40 | 0.01800 / 0.00900 |
| medium | 4 | 8 | 40 | 0.03600 / 0.01800 |
| large | 8 | 16 | 80 | 0.07200 / 0.03600 |
| xlarge | 16 | 32 | 160 | 0.14400 / 0.07200 |
<!-- generated-pricing:end -->
Card-on-file customers: admin-only automatic prepaid refill and hosted card setup; see /docs/prepaid-credit.md. Agents continue MPP.