@powforge/captcha-mcp

Your MCP server returns 429 when agents pound it. captcha-mcp makes them earn their next call instead. Hand the agent a proof-of-work puzzle (free, ~5s of CPU) or a 3-sat Lightning invoice — both are machine-readable backoff signals an autonomous caller can satisfy without an account, email, or API key.

Three tools over stdio or HTTP. Stdlib only. No signup, free fallback, self-hosted, no revenue share.

Why not 429?

429 Too Many Requests is the wrong shape for the agent era. Three patterns recur across MCP server reports:

• Agent frameworks treat 429 as a connection failure. They retry immediately, often with exponential backoff that is still too aggressive, and amplify the overload that triggered the limit in the first place.
• There is no per-caller signal. A 429 fires for the bucket, not the agent. One noisy caller gets every other caller throttled, and the server has no way to ask the noisy one to slow down specifically.
• Retry-After is advisory and frequently ignored. Agents do not consistently parse it, do not consistently respect it, and have no incentive to wait — the cost of retrying is zero.

captcha-mcp replaces the 429 with a 402-style challenge. The next call costs the caller something (CPU seconds or 3 sats). That cost is per-caller, machine-readable, and self-throttling — an agent that cannot solve the puzzle cannot flood the endpoint.

Quickstart

npx -y @powforge/captcha-mcp

No install, no config, no API key. The server starts on stdio and waits for an MCP client.

To wire it into Claude Code, Cursor, or any MCP-compatible host, add to your config:

{
  "mcpServers": {
    "powforge-captcha": {
      "command": "npx",
      "args": ["-y", "@powforge/captcha-mcp"]
    }
  }
}

Or run npx @powforge/captcha-mcp --install to print the config block.

What it does

Wraps the PowForge pow-captcha service (captcha.powforge.dev) as three MCP tools:

The free tier costs the agent ~5-10 seconds of CPU time (SHA-256, default 14 leading zero bits). The paid tier costs 3 sats over Lightning via L402 (RFC 7235 + bolt11 invoice in WWW-Authenticate).

Why this and not OAuth, API keys, or Stripe

Agents do not have email addresses. They do not click confirmation links. They do not enter credit cards. PoW + Lightning is the only auth primitive that works for fully autonomous callers.

Managed MCP auth platforms work, but they charge 100–2000 sats per call on vendor infrastructure — your revenue flows through their rails. This package runs on your server, your Lightning node, your keys. You keep the sats.

Configuration

Set CAPTCHA_URL to point at a different captcha backend. Default is http://localhost:3077 so you can run the full stack locally for development. Production deployments point it at https://captcha.powforge.dev.

CAPTCHA_URL=https://captcha.powforge.dev npx @powforge/captcha-mcp

HTTP Streamable transport

Hosted MCP clients (Smithery, browser-based hosts) need HTTP, not stdio. Pass --http or set HTTP_MODE=1:

HTTP_MODE=1 PORT=3200 npx @powforge/captcha-mcp
# or
npx @powforge/captcha-mcp --http

The server then listens on:

Stateless. No session ids. CORS open (Access-Control-Allow-Origin: *) so browser clients work. Stdio mode is unchanged and remains the default — npx @powforge/captcha-mcp with no flag still talks JSON-RPC over stdin/stdout.

Smoke test the HTTP transport:

HTTP_MODE=1 PORT=3200 node src/server.js &
curl -X POST http://localhost:3200/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1"}}}'

Returns {jsonrpc:"2.0", id:1, result:{protocolVersion:"2024-11-05", capabilities:{tools:{}}, serverInfo:{...}}}.

Local development

Clone the captcha widget repo or run the public service. The MCP server only needs HTTP access to the captcha endpoints listed under status.

git clone https://github.com/zekebuilds-lab/captcha-mcp
cd captcha-mcp
node src/server.js

It prints ready to stderr and waits for JSON-RPC on stdin.

Smoke-test the protocol manually:

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1"}}}' | node src/server.js

You should see a JSON response with serverInfo: { name: "@powforge/captcha-mcp", version: "0.2.5" }.

Token verification from your own backend

When an agent submits a token to your service, verify it without trusting the agent:

curl -X POST https://captcha.powforge.dev/api/token/verify \
  -H "Content-Type: application/json" \
  -d '{"token":"<token-from-verify-tool>"}'

Returns {valid: true, method, issued_at, expires_at} or {valid: false, reason}.

Related packages

• @powforge/captcha — the browser widget for the same service.
• @powforge/mcp-l402-gate — Express middleware to gate any MCP server with L402 + Depth-of-Identity scoring.
• @powforge/mcp-identity — agent reputation oracle. Pair with this gate for first-call abuse protection.

How this compares to other MCP agent-auth primitives

The gate-the-MCP-server space is filling up. Here is the honest landscape, ranked by how directly each tool overlaps with what captcha-mcp does.

PayGated is the closest collision. Same "monetize MCP tools per call" pitch, same self-host + open-source posture, but it settles on Stripe. That means you need a Stripe account in good standing (KYC, a bank, a supported country) to collect, and every caller needs a Stripe customer record before it can pay you a cent. captcha-mcp's differentiator is the no-account path: a non-US agent author pays 3 sats per call in about 200ms with no KYC, or solves a free PoW puzzle if it will not pay at all.

APort and AgentSign sit at a different layer. They record who used a tool under what authority; they do not price the call. They compose with a gate like this one rather than replace it.

None of them price the act of interacting. Every other row assumes the caller is already an authorized identity and meters or audits after that. The PoW tier here is the only mechanism in the table that puts a cost on the interaction itself, not on the identity of the actor. That is the position this package defends.

A longer breakdown against x402-mcp, @agentauth/mcp, and Cloudflare ARC/ACT is at powforge.dev/mcp/compare/x402-mcp.

License

MIT