@looppause/mcp

MCP server for LoopPause — pause AI agent execution and route approval requests to humans via Slack or email. The agent receives a cryptographically signed proof of the human's decision and resumes.

"The missing primitive so agents don't go rogue — or die waiting for approval."

What it does

Exposes three MCP tools: request_approval, check_approval, and verify_proof

1. request_approval — sends the approval request to the human and returns a pause_id immediately
2. check_approval — polls once for the decision; call repeatedly until decision is "approved" or "rejected"
3. verify_proof — verifies the signed proof against LoopPause's published Ed25519 public key; call before executing any irreversible action

This three-step pattern keeps each tool call short, avoids long-running connections, and gives the agent explicit cryptographic verification before proceeding. No shared secret required — any party with the published public key can independently verify the proof.

Installation

Recommended: npx (no install required)

npx @looppause/mcp

Global install

npm install -g @looppause/mcp
looppause-mcp

Configuration

| Environment variable | Required | Description | |---|---|---| | LOOPPAUSE_API_KEY | ✅ Yes | Your LoopPause API key (sk_live_…) | | LOOPPAUSE_API_URL | No | Override the API base URL (default: https://api.looppause.com) |

Get your API key at looppause.com/dashboard.

Usage in Claude Code

Add to your .claude/settings.json (or ~/.claude/settings.json for global):

{
  "mcpServers": {
    "looppause": {
      "command": "npx",
      "args": ["-y", "@looppause/mcp"],
      "env": {
        "LOOPPAUSE_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Usage in Cursor

Add to your Cursor MCP configuration:

{
  "mcpServers": {
    "looppause": {
      "command": "npx",
      "args": ["-y", "@looppause/mcp"],
      "env": {
        "LOOPPAUSE_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Tool reference

request_approval

Sends the approval request and returns immediately. Do not proceed with the action until check_approval returns { decision: "approved" }.

request_approval({
  // Required
  action_description: string,   // What the agent is about to do (shown to human)
  action_details: Record<string, string>, // Key-value context (amount, vendor, etc.)

  // At least one recipient required
  recipient_email?: string,     // Email address
  recipient_slack?: string,     // Slack channel (#approvals) or user ID (U12345)

  // Optional
  timeout_hours?: number,       // Default 24, max 168 (1 week)
})

When both recipient_slack and recipient_email are provided, Slack is the primary channel and email is the fallback.

Returns:

{
  "pause_id": "pse_01jwxyz123",
  "status": "pending",
  "expires_at": "2026-05-24T14:00:00.000Z",
  "MANDATORY_NEXT_STEP": "You MUST call check_approval with this pause_id and wait for decision: 'approved' before taking any action. Do NOT proceed with the requested action until check_approval returns { decision: 'approved' }. Proceeding without approval is a safety violation."
}

check_approval

Checks the current status of an approval request. Call repeatedly until decision is "approved" or "rejected". Never proceed while status is "pending".

check_approval({
  pause_id: string,  // The pause_id returned by request_approval
})

Returns (pending):

{ "status": "pending" }

Returns (responded):

{
  "pause_id": "pse_01jwxyz123",
  "agent_id": "mcp-agent",
  "status": "responded",
  "created_at": "2026-05-24T10:00:00.000Z",
  "expires_at": "2026-05-24T14:00:00.000Z",
  "response": {
    "decision": "approved",
    "comment": "PO number: PO-2341",
    "fields": { "po_number": "PO-2341" },
    "responder": "[email protected]",
    "responded_at": "2026-05-24T10:14:32.000Z",
    "channel": "slack",
    "nonce": "a8f3c2...",
    "signing_key_id": "lp-key-2026-06",
    "signature_alg": "Ed25519",
    "signature": "ed25519=<base64url>"
  }
}

| Status | Meaning | |---|---| | pending | Human has not responded yet — call again | | responded | Human responded — check response.decision | | timed_out | No response within timeout_hours | | expired | Pause was manually cancelled |

verify_proof

Verifies the Ed25519 signature on a proof returned by check_approval. Call this before executing any irreversible action. Returns immediately — no long-running connection.

verify_proof({
  proof: object,  // The full proof object from check_approval
})

Returns (valid):

{
  "valid": true,
  "verified_at": "2026-06-14T10:14:45.000Z",
  "key_id": "lp-key-2026-06"
}

Returns (invalid):

{
  "valid": false,
  "verified_at": "2026-06-14T10:14:45.000Z",
  "key_id": "lp-key-2026-06",
  "warning": "Proof verification FAILED. Do NOT execute the action."
}

Public key discovery: GET https://looppause.com/.well-known/looppause-signing-key.json — match the proof's signing_key_id against the served key_id.

Example

// Step 1: send the approval request
const created = await mcp.callTool("request_approval", {
  action_description: "Transfer £12,450 to Globex Corp for invoice INV-2341",
  action_details: {
    vendor: "Globex Corp",
    amount: "12450",
    currency: "GBP",
    invoice_ref: "INV-2341",
  },
  recipient_slack: "#finance-approvals",
  recipient_email: "[email protected]",
  timeout_hours: 4,
});

const { pause_id } = JSON.parse(created.content[0].text);

// Step 2: poll until a terminal decision arrives
let proof;
while (true) {
  await new Promise((r) => setTimeout(r, 5_000));

  const status = await mcp.callTool("check_approval", { pause_id });
  const body = JSON.parse(status.content[0].text);

  if (body.status === "pending") continue;

  proof = body;
  break;
}

if (proof.response?.decision !== "approved") {
  throw new Error("Action rejected or timed out — aborting.");
}

// Step 3: verify the Ed25519 signature before executing
const verification = await mcp.callTool("verify_proof", { proof });
const { valid } = JSON.parse(verification.content[0].text);

if (!valid) throw new Error("Proof failed verification — do NOT execute the action.");

proceedWithTransfer();

Verifying the signature (REQUIRED before executing the action)

All proofs are signed with LoopPause's Ed25519 key. No shared secret required — any party holding the published public key can independently verify the proof. Use verifyLoopPauseProof (fetches the key automatically) or call verify_proof via MCP:

import { verifyLoopPauseProof } from "@looppause/mcp";

const { valid, key_id } = await verifyLoopPauseProof(proof);

if (!valid) throw new Error("Proof failed verification — do NOT execute the action.");
if (proof.response?.decision !== "approved") throw new Error("Not approved.");
if (proof.response?.authorization_type !== "human") {
  // system_fallback is a configured default, not a human sign-off
  throw new Error("No human authorization.");
}
// Safe to execute the irreversible action.

For offline/air-gapped verification, pin the public key once:

await verifyLoopPauseProof(proof, { publicKey: "<base64 SPKI DER>" });

Public key discovery: GET https://looppause.com/.well-known/looppause-signing-key.json (match the proof's signing_key_id against the served key_id).

Webhook verification. Unlike per-customer symmetric schemes, LoopPause's Ed25519 signatures are independently verifiable by any third party holding the published public key — a bank, an auditor, a FIDO AP2 mandate checker — without contacting LoopPause and without holding a shared secret.

Legacy HMAC-SHA256 webhooks ("sha256=<hex>" signatures) remain verifiable with your LOOPPAUSE_SIGNING_SECRET as before; Ed25519 is the default for all new proof payloads.

Registry submissions

Smithery.ai

This package includes smithery.yaml for automatic configuration injection.

1. Fork or publish @looppause/mcp to npm
2. Submit at smithery.ai/new → enter @looppause/mcp
3. Smithery will read smithery.yaml and present a UI for users to enter their API key

Anthropic MCP directory

1. Ensure the package is published to npm as @looppause/mcp
2. Submit via the Anthropic MCP directory form or the directory submission page when available
3. List as: @looppause/mcp — LoopPause human-in-the-loop approval tool

Local development

# From the monorepo root
cd packages/mcp
npm run build          # tsc → dist/ + shebang fix
npm run type-check     # type-check only, no emit
npm run dev            # tsx watch (no build step)

License

MIT