Build on AgentBee

01 Quickstart

One command wires AgentBee into Claude as an MCP server. After this, your agent can ask for human approval on any critical action, and you can verify the receipt it gets back.

# add the MCP server to Claude Code
claude mcp add agentbee -- npx -y @agentbee/mcp

# no hardware yet? run the simulator — same crypto, sample key
AGENTBEE_MODE=simulator claude mcp add agentbee -- npx -y @agentbee/mcp

Then just ask your agent to do something risky. The bee lights up, you flip APPROVE, the action proceeds with a signature. Ignore it and the action is blocked.

02 The MCP tools

The server exposes two tools. request_approval gates an action behind a human approval; verify_receipt checks a returned receipt against the device public key.

// your agent calls this before doing anything critical
request_approval({
  action: "wire_transfer",
  scope:  "$5,000 to Northwind Ltd",
  risk:   "high"
})
// -> { approved: true, receipt: { fp, sig, ts }, signed: true }
//    or { approved: false } if no one approved in time

verify_receipt({ receipt })   // -> { valid: true, action, ts }

Configure with env: AGENTBEE_MODE (simulator | usb), AGENTBEE_THRESHOLD (graduated trust level that triggers a prompt), AGENTBEE_PORT (serial port for a real device).

03 Auto-gate with a PreToolUse hook

Want approval enforced automatically, without your agent having to remember to ask? Drop in the PreToolUse hook. It classifies risky tool calls — rm -rf, DROP TABLE, git push --force, writes to .env / .ssh, payment and delete MCP tools — and fails closed: no approval, no action.

// .claude/settings.json
{
  "hooks": {
    "PreToolUse": [{
      "matcher": "*",
      "hooks": [{ "type": "command",
        "command": "npx -y @agentbee/hook" }]
    }]
  }
}

Critical calls now pause for your approval on the bee. Everything else runs untouched. The hook is fail-closed by design — if the bee is unreachable, the action is denied, not allowed.

04 The SDK (any agent, any stack)

Not on Claude? The SDK gives you the same gate in plain code — one call, works with any framework or your own loop.

import { AgentBee } from "@agentbee/sdk"

const bee = new AgentBee()           // auto-detects device or simulator

const ok = await bee.approve({
  action: "deploy", scope: "v2.4.0 -> production"
})
if (!ok.approved) throw new Error("blocked by human")

deployToProd()                       // only runs after you approve
await bee.verify(ok.receipt)         // anyone can verify, later

05 What people gate

Real critical actions worth a human decision. Some you approve, some you stop cold.

06 Community

Building something with the bee? Come make noise. Share what you gate, ship integrations, help shape the standard.

Authorization, not authentication. AgentBee approves actions; it does not log you in. It complements YubiKey and your IdP — it does not replace them. Your signing key is generated on the device and never leaves it. We store no keys, no device data, and no email.