# ai2robot-mcp

MCP server that gives any AI agent the ability to dispatch physical tasks to
robot executors and track `task -> proof -> verify -> settle`.

Works with Claude Code, OpenClaw, Cursor, Codex and any MCP-compatible agent.

## What it does

| Tool | Purpose |
|---|---|
| `create_physical_task` | publish a physical task (action, location, proof requirements, budget, deadline) |
| `list_executors` | find executors near a location, filter by action |
| `submit_proof` | submit a proof bundle; deterministic verification gates approval |
| `verification_status` | check status, verdict and escrow id |

## Quick start (local demo mode)

No backend needed. The server runs against an in-memory executor registry and
verifies proof bundles with real SHA-256 / Ed25519 checks.

```bash
npm install
npm run build
npm start
```

Register with your agent:

```bash
# Claude Code
claude mcp add ai2robot -- npx tsx src/index.ts

# OpenClaw
openclaw mcp add ai2robot -- npx tsx src/index.ts
```

Then ask the agent something like:

> "Create a physical task to inspect the store at 31.2304,121.4737, proof =
> photo + GPS trace, budget 10 USDC, deadline in 3 hours. Then show me who was
> matched."

## Example proof bundle

Robots/executors submit a [proof bundle](https://github.com/ai2robot-network/ai2robot-proof)
with a hash-chained action log, GPS traces and an Ed25519 signature. The server
rejects bundles whose chain is broken, timestamps are out of order, or GPS
movement is implausible.

## Roadmap

- Network API mode (`AI2ROBOT_API_URL`) — same tools, live executors
- ERC-8004 executor registration
- VCAP-compatible escrow settlement
- Verifier-node sampling for high-value tasks

## License

MIT