Zero-Trust Policy Enforcement, Upstream Credential Isolation & Per-Call Authorization
The AuthHub MCP Gateway sits between autonomous AI agents and sensitive backend tools. It implements RFC 9728 Protected Resource Metadata, enforces RFC 8707 audience binding to prevent confused-deputy attacks, and evaluates every tool call against a live Google Zanzibar (SpiceDB) relationship graph in under 5 milliseconds — without ever leaking client credentials to upstream servers.
Intercepts tools/call, resources/read, and prompts/get. Evaluates Subject, Action, Resource, and Context in <5ms. Same tool + same subject + different resource = distinct PERMIT vs DENY decision.
Complies strictly with Rule M8. Client tokens authorize the client→gateway leg only. Upstream MCP servers receive short-lived, encrypted credentials injected by the gateway, preventing upstream tool compromise.
Validates that the token's aud or resource indicator matches https://api.authhub.cloud/mcp. Tokens minted for other microservices are immediately rejected with 401.
Consumes Continuous Access Evaluation Profile (CAEP) and Shared Signals (SSF) events. Pushes instantaneous notifications/tools/list_changed over SSE to revoke tool access without terminating active sessions.
Dedicated proxy for database agents (PostgreSQL, Snowflake, BigQuery). Parses SQL ASTs, prevents DROP/TRUNCATE queries, enforces row-level filters, and injects tenant tenancy boundaries.
Returns full protected resource metadata at /.well-known/oauth-protected-resource. Standard 401 challenge includes the metadata pointer, enabling automated discovery in any compliant MCP client.
Click each step to inspect the exact wire protocol headers, payloads, and policy evaluations.
The client (e.g. MCP Inspector, Claude Desktop) connects without authorization credentials.
POST /mcp HTTP/1.1
Host: api.authhub.cloud
Content-Type: application/json
Accept: application/json, text/event-stream
{
"jsonrpc": "2.0",
"id": 1,
"method": "initialize",
"params": {
"protocolVersion": "2025-11-25",
"capabilities": { "roots": { "listChanged": true } },
"clientInfo": { "name": "mcp-inspector", "version": "0.14.1" }
}
}Empirical testing records from the OpenID MCP security interoperability programme.
Verified with official Anthropic inspection harness across all gated operations (tools/call, resources/read, prompts/get).
Validated RFC 9728 resource discovery and automated client metadata document registration without static pre-shared keys.
Verified cross-language MCP client connectivity, OAuth 2.1 token issuance, and SSE session management.
Demonstrated automated enterprise governance lifecycle: role changes trigger immediate tool revocation over SSF/CAEP.
Deploy AuthHub MCP Gateway in front of internal tools or cloud databases in minutes.