waw-mcp — World Agentic Web MCP server

Automatic authentication & identity for AI agents over a single stdio MCP connection — so an agent creates its own accounts on third-party services, autonomously. The MCP server of WAW. It gives an agent:

• a real, valid email address as its identity (mailbox_provision),

• the ability to read OTP / verification codes sent to it (otp_get),

• autonomous signup to a website with a real browser (signup_run),

• page perception combining vision + UI tree (browser_open / browser_observe …),

• and discovery / publishing of MCP servers by intent (registry_search / registry_register).

This package is the standalone, runs-anywhere distribution: one bundled file, started by node over stdio, configured only by environment variables (no secret ever lives in the repo). The full source, the HTTP/SMTP gateway that receives email, and the Docker demo live in the main repo: https://github.com/Revens2/waw.

Validated end-to-end: an agent connects to this published package, provisions an owner-linked identity, signs up on a live site, receives the OTP via a signed webhook, and the account is created — with no human typing a single field. See Autonomous workflow below.

Install

npm install -g waw-mcp        # global `waw-mcp` command
# or run without installing:
npx waw-mcp

better-sqlite3 (native) is installed automatically. For signup_run and the browser_* tools you also need a browser once: npx playwright install chromium.

Related MCP server: useblip/email

Configure (environment variables — never commit secrets)

Use from an MCP client (Claude Code, etc.)

Add to your MCP config (e.g. .mcp.json or claude mcp add). Secrets go in the env block, which lives in your local client config — not in any repo:

{
  "mcpServers": {
    "waw": {
      "command": "waw-mcp",
      "env": {
        "VAULT_PASSPHRASE": "your-long-random-passphrase",
        "MAIL_DOMAIN": "agents.example.com"
      }
    }
  }
}

# CLI equivalent
claude mcp add waw --env VAULT_PASSPHRASE=… --env MAIL_DOMAIN=agents.example.com -- waw-mcp

Tools (16)

Browser realism: the Playwright drivers apply generic hardening (no AutomationControlled flag, navigator.webdriver masked, realistic viewport) — WAW_STEALTH=0 to disable. This is not a bypass for commercial anti-bot (Turnstile/Akamai/PerimeterX); those stay human-in-the-loop. Respect each site's ToS.

Autonomous workflow (the point)

An agent authenticates itself and creates an account with zero human form-filling:

1. mailbox_list                         → empty? run onboarding
2. mailbox_provision { ownerEmail }      → identity: you.agent@MAIL_DOMAIN
3. signup_run { targetUrl, email, … }    → fills the form with a vaulted password,
                                           waits for the OTP, submits it
   └─ OTP arrives via a signed webhook into the shared WAW DB, read back automatically
4. account created ✅   (resumable & idempotent — safe to retry)
5. registry_register { … }               → publish the service so other agents find it

Hard sites where fixed selectors break? Swap step 3 for the perception loop: browser_open → look at the screenshot + element refs → browser_fill/browser_click by ref → browser_observe to verify → repeat.

First connection — link a human owner

On a fresh connection the agent calls mailbox_list; if it is empty, it asks the human operator for their real email and provisions its primary identity with ownerEmail set to it. The agent address is then derived from the owner's, and the ownership is recorded — every autonomous identity traces back to a real human (accountability + recovery):

owner you@example.com  →  agent you.agent@MAIL_DOMAIN

The owner email is supplied at runtime by the human; it is never stored in this repo.

Receiving real email

This server reads OTPs from the WAW database. To make real messages land there, pair it with the WAW gateway (main repo) sharing the same DB_PATH: the gateway exposes a signed inbound webhook (Mailgun / Postmark / generic) and an optional self-hosted SMTP server. See https://github.com/Revens2/waw for the gateway, the security model, and a Dockerized end-to-end demo.

Security

Secrets come only from environment variables; .env is git-ignored and .env.example carries placeholders. Vault data is AES-256-GCM encrypted at rest, OTP codes are never returned over the wire, and logs redact sensitive fields. Full threat model in the main repo's SECURITY.md.

License

MIT