MCP servers

An MCP server is a project type for Model Context Protocol endpoints. Agents must authenticate before they can call a tool. Create one from Projects by choosing MCP server.

MCP Auth appears in the sidebar only for that project type.

Three MCP meanings in Authdog: an MCP project (this doc) is your customer’s tool server and OAuth resource; the Agents module is the machine-identity registry and trust store for non-human callers; the Management API MCP is Authdog’s own control-plane tools for operators — not your product’s MCP server.

Settings

Open MCP Auth → Settings to enable MCP Auth for the environment, set the default resource (RFC 8707), and turn on strict resource mode. See MCP authentication for the authorize and token flow.

Playground

Open MCP Auth → Playground. Connect with SSO to mint a short-lived bearer for the environment, then import tools. Tokens are scoped to the selected environment — use dev while iterating.

A successful playground call proves this console session can invoke that tool on that server. It does not authorize every agent identity in the trust store. Register those in Agents (sales-gated).

After registration

Each project can hold several environments, so staging tools never resolve in production.

Controls in `@authdog/mcp-sdk`

These run inside your server. Each is opt-in.

• Trust store — verified non-human identities from Agents. Issuer revoke blocks new tokens when the subject is the OIDC client_id. Outstanding JWTs stay valid until expiry unless this control is in enforce.
• Per-tool authorization — AuthZEN check before the tool runs.
• Audit — invocation, denial, failure. No raw arguments.
• Threat detection — Lidar signals for poisoning, injection, exfil, rate abuse.

Happy path is enforce. Use off / monitor only while staging. See MCP security and machine identity.

@authdog/mcp-sdk is source-only in the Authdog monorepo (packages/mcp-sdk). It is not published on npm.

Related