MCP Identity Gateway

The MCP Identity Gateway sits between AI agents and MCP servers, enforcing policies and injecting per-user credentials.

Two-Policy Model

The Gateway requires two Access Policies:

Policy 1: Client-to-Gateway

Client Workload — AI agent identified by redirect URI
Server Workload — The Gateway endpoint
Trust Provider — OIDC token validating the human user
Credential Provider — OIDC token embedding user identity

Policy 2: Gateway-to-Server

Client Workload — The Gateway itself
Server Workload — The actual MCP server
Trust Provider — Validates the Gateway’s token
Credential Provider — Per-user OAuth token for the MCP server

Per-User Credential Isolation

Each user gets their own MCP server credentials:

User authenticates via IdP when connecting through the Gateway
First-time access triggers OAuth consent with the MCP server
Dipole stores each user’s token separately
Subsequent requests inject that user’s specific token
Audit trail shows exactly which user accessed which tool