MCP SDK

Authdog ships two MCP server SDKs. They solve different problems and are not interchangeable, so pick by what your MCP server needs to enforce.

Token verifier SDKs

An MCP client obtains an Authdog access token, for example through MCP authentication, and sends it as a bearer token. AuthdogVerifier calls the environment's GET /v1/userinfo with that token and returns an AuthInfo. The Node.js verifier can instead verify a JWT locally against your environment's JWKS when you set issuer, which also fills in scopes and expiry from the token. The list below describes the userinfo path that every language shares:

• subject, username, and email for the authenticated user
• expiresAt, derived from the remaining session lifetime
• scopes, currently always empty because userinfo does not return scopes yet
• token and the full user object

A 401 from Authdog raises an authentication error your transport should turn into an HTTP 401. Any other failure raises a separate API error, so you can tell a rejected token from an unreachable Authdog.

Pick your language:

MCP runtime guard

The TypeScript runtime SDK runs inside your MCP worker and layers four independent controls over each tool call. Adopt them one at a time:

1. Authenticate: verify the caller's access token against your environment's JWKS, with RFC 9728 protected-resource metadata.
2. Trust store: require the calling identity to be a verified non-human identity within its allowed scopes.
3. Authorize: ask the AuthZEN decision point whether this identity may invoke this specific tool.
4. Telemetry: emit invoked, denied, and failed events for audit and SIEM.

Trust store and authorization default to non-breaking modes (off and monitor). Move each to enforce after you have watched the decisions in the console. The guard is private and is not on npm, so install it from a clone of the monorepo. Use the verifier above if you only need to authenticate callers.

import {
  authenticateRequest,
  checkTrustStore,
  createAuthdogTelemetry,
  withAuthdogToolGuard,
} from "@authdog/mcp-sdk"

const auth = await authenticateRequest(request, env)
if (!auth.ok) return new Response(auth.message, { status: auth.status })

const trust = await checkTrustStore(auth.claims, env)
if (!trust.ok) return new Response(trust.message, { status: trust.status })

const telemetry = createAuthdogTelemetry(env)
const readFile = withAuthdogToolGuard("my-server", "read_file", env, handler, {
  telemetry,
})

Configure it with these variables on your worker:

AUTHDOG_DEV_MODE=true allows http://localhost issuers for local development. Never set it in production.

To detect a server changing its tools after registration, report a tool manifest hash with computeToolManifestHash and pin it on the trust-store entry.

Related

• MCP authentication
• MCP OAuth vs platform
• Console MCP module
• Machine identity