Introduction to MCP

The Discover MCP server is a remote Model Context Protocol service. It lets AI coding clients — Claude Code, VS Code, Cursor, Kiro, Windsurf, claude.ai — query your Discover knowledge bases (JIRA, Confluence, Codebase, Database, API, Documents etc) over a single HTTPS endpoint.

Your assistant gets the same grounded, citation-bearing retrieval that Discover agents use, without leaving your editor.

Endpoint	https://<your-discover-host>/mcp
Transport	Streamable HTTP (stateless, JSON responses)
Auth	OAuth 2.1 — Dynamic Client Registration + PKCE + loopback redirect
Scopes	mcp:tools offline_access

Getting your endpoint

The endpoint is specific to your Discover deployment. Ask your administrator for the host — the path is always /mcp. Examples on these pages use https://<your-discover-host>/mcp as a placeholder.

---

How authentication works

The server is an OAuth 2.1 resource server. You never paste an API key. On first connect your client:

1. Hits /mcp and gets a 401 pointing at the Protected Resource Metadata.
2. Discovers the authorization server and registers itself dynamically (DCR).
3. Opens your browser to the login page.
4. You sign in — via your organisation’s SSO, or a username and password if your identity provider issues one.
5. The browser redirects to a local localhost:<port>/callback. Your client stores the access token plus a refresh token (from offline_access).

Tokens refresh automatically, so you log in once. All you need is an account in your deployment’s identity realm — there are no per-client secrets.

---

Two ways to connect

Mode	When to use it
Native remote HTTP + OAuth	Claude Code, VS Code 1.102+, recent Cursor. Give the client the URL and it handles the rest.
mcp-remote bridge	Any stdio-only client, or one without built-in OAuth. A small local proxy (npx mcp-remote) runs the OAuth flow and bridges stdio ↔ HTTP. Universal fallback.

→ Setup for each client: Connecting a Client

Once connected, call health to see the tool surface your deployment exposes.