Many vendors stop at **OAuth for MCP**: metadata, consent, and a bearer token. That
proves a client got scopes; it does not prove a specific **tool** should run, that the
caller is a **verified agent**, or that you can **audit and revoke** machine access
like human sessions.

Authdog’s MCP Auth platform keeps OAuth 2.1 and RFC 9728 as the front door, then adds
the controls operators expect in production.

## Comparison

| Capability | OAuth-only MCP connect | Authdog MCP Auth platform |
|------------|------------------------|---------------------------|
| Protected-resource metadata (RFC 9728) | Yes | Yes — per environment issuer |
| PKCE authorize + token | Yes | Yes — hosted consent, DCR/CIMD options |
| Scope strings on token | Yes | Yes — catalog + admin-only consent flags |
| **Per-tool** allow/deny after token | No — app must build | Yes — AuthZEN `MCP_TOOL` policies |
| Verified agent / trust store | No | Yes — [Agents](/docs/console/agents) + enforce mode |
| Runtime audit (`MCP_TOOL_DENIED`, …) | Rarely | Yes — `@authdog/mcp-sdk` telemetry → Audit/SIEM |
| User revoke from Connected Apps | Varies | Yes — consent + token revocation |
| Enterprise SSO on same issuer | Varies | Yes — same environment IdP as hosted auth |
| SCIM deprovision → revoke grants | Varies | Yes — refresh blocked after deactivation |
| Hosted edge gateway | Varies | Optional — `{label}.mcp.authdog.com` (024) |
| Management MCP for operators | N/A | `https://api.authdog.com/mcp` (control plane) |

## When OAuth alone is enough

Use a minimal OAuth layer if you only need “signed-in user can hit any tool the server
exposes” and you trust your server code to enforce everything else. That fits early
prototypes and internal tools with no agent fleet.

## When you want the full stack

Choose Authdog MCP Auth when:

- Non-human **agents** call tools and need verification, pinning, and rug-pull hooks.
- Security wants **deny with evidence** (`MCP_TOOL_DENIED` in audit/SIEM), not silent 403s in logs.
- Compliance wants **revocation** tied to directory lifecycle (SCIM, deactivate user).
- Product teams want **scope + policy** separation: OAuth for consent, AuthZEN for tools.

## Get started

1. Create an **MCP server** project and enable **MCP Auth → Settings**.
2. Integrate `@authdog/mcp-sdk` with trust store + AuthZEN enforce ([MCP security](/guides/mcp-security)).
3. Forward `mcp_runtime` events to your SIEM ([MCP authentication → Audit & SIEM](/docs/mcp-authentication#audit-and-siem)).

## Related

- [MCP authentication](/docs/mcp-authentication)
- [Machine identity](/docs/machine-identity)
- [Authorization (AuthZEN)](/docs/authzen)