# Vercel Sandbox Stack

In the [AI Enablement Stack](/docs/introduction/ai-enablement-stack), **Blekline is Layer 4** and **Vercel Sandbox is Layer 1** (Firecracker microVMs for untrusted / AI-generated code).

This is **not** Blekline hosting on Vercel — it is running **agent code execution** in [Vercel Sandbox](https://vercel.com/docs/sandbox) while Blekline governs MCP tool calls at the ingress.

| Layer | Product | Role |
|-------|---------|------|
| **L4** | Blekline | Mask prompts, classify risk, govern MCP tool calls, audit |
| **L1** | Vercel Sandbox | Isolated microVM, Node/Python runtimes, ms cold start |

## Wiring

1. Connect `@blekline/mcp-proxy` — [MCP proxy](/docs/mcp/proxy)
2. Use `@vercel/sandbox` in your agent app for code execution
3. Optionally connect [Vercel MCP](https://mcp.vercel.com) for deploy/project ops (separate from sandbox execution)

```bash
export BLEKLINE_DOWNSTREAM_SERVER=vercel
export BLEKLINE_WORKSPACE_TOKEN=blw_...
# Vercel Sandbox auth via OIDC in CI or access token locally
```

**Flow:** L5 Agent → L4 Blekline proxy → your agent orchestrator → L1 Vercel Sandbox microVM.

See [Vercel Sandbox docs](https://vercel.com/docs/sandbox) and [durable AI agent guide](https://vercel.com/kb/guide/how-to-build-a-durable-ai-code-agent-on-vercel).

---

**Next steps:** [Sandbox providers](/docs/integrations/sandbox-providers) · [AI Enablement Stack](/docs/introduction/ai-enablement-stack) · [MCP proxy](/docs/mcp/proxy) · [Open workspace](https://app.blekline.com)