Protect an application

This section takes one MCP server from unprotected to launched: every tool call requires a valid OAuth token with the right scopes, and admins control access from the dashboard without touching your code. Budget ~15 minutes for a first pass.

Before you start

• A workspace on https://mcpauthz.com
• An MCP server (or any HTTP API) with a public URL — for local development, an ngrok tunnel works (local development)

The path

Each step is one page, one dashboard tab, one decision:

Step	Page	You decide
1	Register an application	Name, public URL, protected path, scope vocabulary
2	Set up the environment and SDK	Credentials, env vars, and the mount_mcp wiring
3	Run the protection test	Eight checks that show exactly where your setup stands
4	Tool inventory	How AuthSec learns your tool list (manifest or scan)
5	Define scopes	The permission vocabulary
6	Map tools to scopes	Which scope gates each tool — deny by default
7	Access policy & default role	Who gets which role, and what a new caller receives
8	Launch	Flip enforcement on
9	Clients and connections	Who may call: DCR, CIMD, pre-registered
10	Monitor	Watch drift and runtime after launch

In a hurry? From zero to launched is the same path condensed to one page. Prefer code-first? Protect your MCP server (Python) walks the identical flow from the SDK side.

What you'll have at the end

• Unauthenticated calls → 401 with a WWW-Authenticate challenge that tells agents where to authenticate
• Valid token, missing scope → 403 insufficient_scope for that tool
• Valid token, granted scope → the tool runs — and tools/list is filtered to what the caller may see
• Policy changes propagate to your running server in ~30 seconds, no redeploy

Then: nothing can call it yet

Launching enforces policy — it doesn't hand out tokens. The next section onboards your callers:

→ Connect agents and services — service accounts (client secret, private-key JWT, Kubernetes/SPIFFE) and AI agents acting on behalf of users (ID-JAG).