Cortex Gateway

Cortex Gateway

A federated MCP gateway: one spec-compliant, OAuth-protected MCP server in front of N plain-HTTP backends.

Your business apps stay ordinary web services. Each one exposes a single POST /api/cortex/backend endpoint (a ~120-line contract, no MCP library, no stdio). The gateway discovers their tools, merges them into one MCP catalog, enforces OAuth 2.1 + scopes, routes tools/call to the owning backend, and keeps a pseudonymized audit trail.

Built for one requirement: everyone in a company should be able to hand their own agent the keys — with exactly that person's rights, nothing more, nothing less. And when agents become autonomous assistants, the answer stays the same: the agent borrows the person's identity, the apps keep enforcing that person's rights. The gateway decides nothing, so there is nothing new to trust.

Put another way: zero-trust principles for AI agents — the missing link between your IAM (who your users are, what they may do) and the MCP ecosystem (how agents call tools). Not a full ZTNA product; the identity-and-access layer for agents.

A backend is a dedicated MCP reduced to its essence: a tool catalog plus tool invocation (and optional prompts/resources) over bare HTTP JSON-RPC — the transport and lifecycle machinery (initialize, sessions, SSE, version negotiation) lives once, in the gateway. Because the contract is a semantic subset of MCP, a native MCP server can also be federated through the built-in MCP→backend proxy adapter (docs/mcp-adapter.md).

[MCP agent: Claude Desktop / claude.ai Custom Connector / any MCP client]

│ HTTPS + OAuth 2.1 JWT (Bearer) ▼ [cortex-gateway] ←— thin gateway, no business logic │ HTTPS + the same JWT propagated (RFC 8707) ▼ [your backends] ←— domain owners, plain HTTP, own their ACLs

Why

• Permissions match the underlying app, automatically, at the user level. The gateway never copies or mirrors permissions — it propagates the real user identity (your JWT to first-party backends, the user's own linked token to proxied third-party MCP servers), so each app's native permission model applies per user, with nothing to sync and no service-account flattening. Unlike hosted tool aggregators, the token vault and the audit trail stay on your infrastructure.
• One perimeter, not N connectors. Wiring each app as its own MCP connector also keeps native permissions — and leaves you with N consents, N token stores, no shared audit, no cross-app entitlements, and a flooded tool list. The gateway collapses that to one OAuth surface without giving up the per-user model: aggregation without permission loss.
• Zero MCP lock-in in your apps. Backends speak a minimal JSON-RPC contract over plain HTTP. Remove the gateway and you can still call them directly (tests, batch jobs, other integrations).
• One JWT, N backends. The agent authenticates once; the gateway propagates the token; every backend re-validates it and applies its own permissions. Revocation at the authorization server cuts everything.
• Failure isolation. An unreachable backend just disappears from tools/list; the rest keeps working.
• Context-aware federation. Backend filtering and a compact "search" mode keep the tool catalog from flooding the agent's context (docs/tool-search-mode.md).
• Agent feedback loop. Agents can file report_missing_capability tickets when a tool is missing or insufficient — deduplicated, triaged, optionally pushed to a webhook when blocking.

Use cases

Company-wide agent surface. An organization runs N internal apps (CRM, quality docs, billing, analytics...). Each app adds the ~120-line backend endpoint; the gateway exposes them as ONE MCP connector protected by the company's SSO. Employees plug a single URL into Claude Desktop / claude.ai and get exactly the tools their token scopes allow, with a central audit trail. This is the setup the gateway was born in.

Product builder. You ship several products and want agents (yours or your customers') to operate them. Instead of maintaining one MCP server per product, every product implements the backend contract and the gateway is your single, versioned, OAuth-protected agent API. Adding a product to the agent surface is one env var.

Thematic hub / curated registry. Run a gateway as a *topic* endpoint — e.g. "all open-data tools for domain X" — that federates several providers behind one URL with one token. The scope model gives you per-provider opt-in, get_help/get_snapshot give agents self-describing discovery, and the audit trail tells you what is actually used. Providers either speak the (deliberately tiny) backend contract natively, or — for off-the-shelf MCP servers — get fronted by the built-in MCP→backend proxy adapter (docs/mcp-adapter.md).

Free / paid tool tiers. Scopes are entitlements. Let your authorization server grant mcp:yourapp:basic to free users and mcp:yourapp:pro to paying ones (your billing webhook updates the grant): the gateway then shows and allows each caller exactly the tools of their plan — no paywall logic in the gateway or the backends, tools just declare their scope. Revocation and downgrades propagate through the normal OAuth chain.

Federating native MCP servers (adapter, beta)

The built-in MCP→backend proxy adapter lets a bundle mix contract backends and off-the-shelf native MCP servers (Canva, Figma, ...): the adapter is an MCP client downstream (initialize, sessions, SSE framing) and a plain backend upstream, so the gateway core does not change. Per-user downstream OAuth is handled by a token vault (AES-256-GCM at rest) and a linking flow (RFC 9728 discovery, Dynamic Client Registration, PKCE): each user consents once per provider, then agents are identified on the whole bundle with a single Cortex token. See docs/mcp-adapter.md for a worked "design bundle" example (Canva + Figma + your own backend).

Roadmap

• Per-backend audiences — exchange the caller's token for one audienced to the target backend (RFC 8693), so a backend can no longer replay what it receives against the gateway (docs/security.md §6).
• Tool annotations and provenance — carry readOnlyHint / destructiveHint from the backend contract into tools/list, mark results with their origin, and confirm writes.
• 2026-07-28 completeness — subscriptions/listen for list-change notifications to stateless clients; backends are still called over the backend contract, unaffected by the client's protocol generation.
• Machine identity for discovery — replace the static technical token with a client_credentials flow once your AS supports it.
• Shared event bus / rate-limit store — for multi-instance deployments.

Features

• MCP spec 2026-07-28 (stateless: server/discover, no session, Mcp-Method / Mcp-Name headers validated against the body, cache hints) and 2025-11-25 / 2025-06-18 (Streamable HTTP: initialize, GET SSE for listChanged notifications, DELETE session termination) on the same endpoint, chosen per request
• Primitives: tools, resources (incl. URI templates), prompts
• OAuth 2.1 resource server: JWKS verification, audience per resource (RFC 8707), protected-resource metadata (RFC 9728), optional RFC 7662 introspection for revocation
• Scope-based visibility: agents only see (and can call) the tools their token scopes allow; scopes_supported in discovery is derived live from the federated catalog
• Dynamic discovery: backends are polled every 60s; new tools appear without redeploying the gateway, with SSE tools/list_changed push
• Builtins: whoami (aggregated identity across backends), find_tools, report_missing_capability, list_cortex_tickets, list_cortex_resources, read_cortex_resource, plus a self-describing cortex://architecture resource generated live
• Audit trail: one JSON line per call on stdout (hashed email, hashed params) + optional PostgreSQL persistence with retention cron
• Origin allow-list (anti DNS-rebinding), per-token rate limiting, optional pool sandbox
• Checked against the official MCP conformance suite in CI, through a real OAuth path: 22 of 32 server scenarios pass; the other 10 exercise optional capabilities the gateway does not declare (completions, logging, subscriptions, progress, sampling, elicitation) — listed with the reason in scripts/conformance/baseline.yml

Quickstart (no OAuth server needed)

git clone https://github.com/wellknownmcp/cortex-gateway

cd cortex-gateway npm install

1. Start the demo backend (dependency-free)

node examples/demo-backend/server.mjs &

2. Configure the gateway

cat > .env.local <<'EOF' OAUTH_ISSUER=https://auth.example.com CORTEX_BACKENDS=demo CORTEX_BACKEND_DEMO_URL=http://127.0.0.1:4820 CORTEX_TECHNICAL_TOKEN=demo-technical-token CORTEX_DEV_BYPASS_TOKEN=dev-secret CORTEX_DEV_BYPASS_SCOPES=mcp:demo:read EOF

3. Run it

npm run dev

4. Talk MCP (dev bypass replaces the Bearer JWT locally)

curl -s http://localhost:3213/mcp \ -H 'Content-Type: application/json' \ -H 'X-Dev-Mode: dev-secret' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq '.result.tools[].name'

→ whoami, ..., demo_get_help, demo_echo, demo_get_time

curl -s http://localhost:3213/mcp \ -H 'Content-Type: application/json' \ -H 'X-Dev-Mode: dev-secret' \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"demo_echo","arguments":{"message":"hello"}}}' | jq

In production you point OAUTH_ISSUER at your OAuth 2.1 authorization server (any server that issues RS256 JWTs with a JWKS endpoint and supports the scope claim), and MCP clients connect to https://your-host/mcp with a Bearer token whose aud is the gateway's canonical URI.

stdio bridge (for stdio-only clients and directory sandboxes)

The gateway is a remote Streamable HTTP server, but some MCP clients — and the Docker inspection harnesses of directories like Glama — only speak the stdio transport. scripts/stdio-bridge.mjs bridges the two: it boots the production build against an ephemeral local OAuth issuer (the real JWT verification path, no bypass), mints itself a short-lived token, and relays newline-delimited JSON-RPC between stdin/stdout and POST /mcp.

npm run build

npm run start:stdio # stdio MCP server on stdin/stdout

Self-granted scopes come from BRIDGE_SCOPES (gateway builtins need none); the gateway port from BRIDGE_GATEWAY_PORT (default 3213). Design notes and the gotchas (stdout purity, protocol-version negotiation, session echo): https://cortex-gateway.dev/guides/expose-http-mcp-server-over-stdio/.

Adding a backend

1. Implement the contract in your app — one POST endpoint, list_tools + your tools (docs/backend-contract.md, reference implementation in examples/demo-backend/).
2. Declare it:

   CORTEX_BACKENDS=demo,docs

CORTEX_BACKEND_DOCS_URL=http://127.0.0.1:4001

3. Done. The gateway discovers `docs_*` tools within 60s and pushes

tools/list_changed to connected clients.

Configuration