Cortex Gateway
This server exposes a federated, OAuth-protected MCP gateway that lets an agent discover and call permitted tools across backends, inspect identity/resources, and file capability tickets.
Identify the caller and aggregate backend identity/roles/scopes via
whoami.Discover and call federated backend tools, filtered by the caller’s OAuth scopes.
List and read gateway/backend MCP resources, including
cortex://architecture.File missing-capability or insufficient-tool reports with
report_missing_capability.Review your own previously filed tickets with
list_cortex_tickets.Use prompts/resources and relay MCP Apps interactive views where declared.
Rely on OAuth 2.1 JWT verification, per-user token forwarding, audit logging, tool-integrity checks, and failure isolation.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Cortex Gateway show me the tools I have access to"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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 ACLsWhy
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_capabilitytickets when a tool is missing or insufficient — deduplicated, triaged, optionally pushed to a webhook when blocking.
Related MCP server: Multi-MCP Hub
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/destructiveHintfrom the backend contract intotools/list, mark results with their origin, and confirm writes.2026-07-28 completeness —
subscriptions/listenfor 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_credentialsflow 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-Nameheaders validated against the body, cache hints) and 2025-11-25 / 2025-06-18 (Streamable HTTP:initialize, GET SSE forlistChangednotifications, DELETE session termination) on the same endpoint, chosen per requestPrimitives: tools, resources (incl. URI templates), prompts
MCP Apps (
io.modelcontextprotocol/ui): the interactive views your backends declare are relayed —_meta.ui, theui://resource,structuredContent— under the scope of the tool that declares them, for contract backends and native MCP servers alike (docs/mcp-apps.md)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_supportedin discovery is derived live from the federated catalogDynamic discovery: backends are polled every 60s; new tools appear without redeploying the gateway, with SSE
tools/list_changedpushBuiltins:
whoami(aggregated identity across backends),find_tools,report_missing_capability,list_cortex_tickets,list_cortex_resources,read_cortex_resource, plus a self-describingcortex://architectureresource generated liveAudit 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"}}}' | jqIn 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/stdoutSelf-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
Implement the contract in your app — one POST endpoint,
list_tools+ your tools (docs/backend-contract.md, reference implementation inexamples/demo-backend/).Declare it:
CORTEX_BACKENDS=demo,docs CORTEX_BACKEND_DOCS_URL=http://127.0.0.1:4001Done. The gateway discovers
docs_*tools within 60s and pushestools/list_changedto connected clients.
Configuration
Everything is env-driven — see .env.example for the full annotated list. The essentials:
Variable | Required | Purpose |
| prod | Canonical MCP resource URI (RFC 9728), default JWT audience |
| yes | Your OAuth 2.1 authorization server |
| yes | Federated backends |
| yes | Static token for catalog discovery (catalog methods only) |
| prod | Web origins allowed (exact or |
| no |
|
| with | Secret for |
| with | Where approved tool definitions persist; without it a restart re-approves the current state |
| no | Ed25519 signing of that store. Public key alone = the gateway verifies approvals but cannot mint them |
| no | Baseline scope demanded before any dispatch |
| no | PostgreSQL for audit persistence + gateway tickets |
| no | Webhook for blocking missing-capability tickets |
| no | Scope prefixes the backend may declare (default |
| no | URI schemes the backend may claim for resources (default |
| behind a proxy | Number of trusted proxies in front, for reading |
| no |
|
Security model
Two paths, and the difference between them is the whole design: the catalog is discovered in the background with a technical token that can only list, and a call carries the user's own token all the way to the backend.
sequenceDiagram
autonumber
participant Agent as AI agent
participant GW as Cortex Gateway
participant AS as Authorization server
participant App as Backend app
rect rgba(128,128,128,0.10)
note over GW,App: Catalog refresh — every 60s, no user involved
GW->>App: tools/list (CORTEX_TECHNICAL_TOKEN)
App-->>GW: full tool catalog
note over GW: Fingerprint each definition<br/>(rug-pull detection)
end
note over GW,AS: JWKS fetched once and cached —<br/>tokens are verified locally, not per call
Agent->>GW: tools/list (user JWT)
note over GW: Verify signature, issuer, audience<br/>Filter the cached catalog by the token's scopes
GW-->>Agent: only the tools this user may call
Agent->>GW: tools/call demo_echo (user JWT)
note over GW: Check the scope declared by that tool
GW->>App: tools/call — the user's JWT, forwarded
note over App: Applies its own roles and ACLs<br/>to that user
App-->>GW: result, under the user's own rights
GW-->>Agent: result
note over GW: One audit line: who, which tool,<br/>which backend, which scopeThe agent never holds a credential the user does not have, and no over-privileged service account exists anywhere on the path.
Agent access is secured by construction, not by gateway policy. The properties below are what to demand from any MCP tooling you wire into an AI app:
OAuth 2.1, not shared API keys — every caller authenticates as themselves; tokens are per-user, scoped and revocable.
No permission flattening — the real user's identity is propagated to each backend, so no over-privileged service account exists and the agent gets exactly the user's own rights.
Least privilege — the tool catalog is scope-filtered per caller; agents only see the tools their token allows.
Verifiable — the whole OAuth discovery chain is walkable without a token, so you (or a third-party scanner) can confirm the posture before connecting.
Under the hood:
The gateway decides nothing about business permissions. OAuth scope is the front door (checked twice: gateway + backend); application roles and ACLs live in each backend.
The static technical token can only reach catalog methods (src/contract/static-token.ts); every data method requires the end user's JWT.
Audit is pseudonymized by design (hashed email, hashed params). Pseudonymized is not anonymous: hashed identifiers remain personal data under GDPR Art. 4(5), so the audit trail stays in your record of processing and needs a retention period.
Sessions are bound to the token's
sub; foreign session ids get 404.
Controls a federating gateway can enforce that a single server cannot. Because it sees every backend's tool definitions and every hop to them, the gateway is the place to catch what the ecosystem's incident reports keep finding — full details in docs/security.md:
Rug-pull detection. Tool definitions are fingerprinted (
description,inputSchema,scope,version) at first sight and re-checked at every refresh. A backend that rewrites what a tool claims to do while keeping its name is reported, and withCORTEX_TOOL_INTEGRITY_MODE=blockthe tool is quarantined until an operator reviews it over/api/admin/tool-integrity— an HTTP endpoint, not an MCP tool, so a model cannot clear a rug pull on its own. Name-level change detection — what most implementations do — misses exactly this attack.No plaintext to remote hosts. Every federated call forwards the caller's token, so an
http://backend URL pointing anywhere but loopback is refused at load. Same validation on the OAuth endpoints the adapter discovers from a third-party server's metadata, which is the class of bug behind CVE-2025-6514.Attributable audit. Each line records which backend served the call and under which scope — not just the tool name.
Backends stay in their lane. A backend may only declare scopes under its own prefix and claim its own URI schemes; prompts are prefixed like tools. Resource reads are checked against the scope of the template the URI falls under, not only filtered at listing.
Honest scope: approvals persist across restarts (CORTEX_TOOL_BASELINE_FILE)
and can be operator-signed, so this proves "this definition changed since you
approved it here", not "this is the definition the vendor published".
Token forwarding, stated plainly. The caller's token is forwarded to the backend that serves the call, and its audience is the gateway. That is what removes the service account, and it means a federated backend is inside the trust boundary of the tokens it receives: one that is compromised could replay a token against the gateway. The gateway limits who receives it (only the serving backend, only backends the caller holds a scope on, never over plaintext); per-backend audiences through token exchange are on the roadmap. Details and recommendations: docs/security.md §6.
Compliance. These are the controls audits test for on automated access: least-privilege scopes, per-user identity (no over-privileged service account), a per-call attributable audit trail, and central revocation. They map to ISO 27001:2022 Annex A access-control and logging controls (A.5.15, A.5.16, A.5.17, A.5.18, A.8.2, A.8.15, A.8.16) and to SOC 2 CC6/CC7 — both of which apply today. The EU AI Act articles usually cited (Art. 12 record-keeping, Art. 14 human oversight) govern high-risk AI systems only, and after the 2026 Digital Omnibus those obligations were deferred to 2 Dec 2027 (stand-alone Annex III) and 2 Aug 2028 (Annex I embedded); most internal agent deployments are not high-risk, and a gateway is access infrastructure rather than an AI system. Self-hosted means the audit trail and token vault stay in your perimeter — no extra sub-processor under GDPR Art. 28, none in your SOC 2 scope. Cortex supplies the controls, not a certification. Full mapping: https://cortex-gateway.dev/answers/ai-agent-compliance-controls/
Development
npm run typecheck # tsc --noEmit
npm test # vitest
npm run build # prisma generate + next buildThe database is optional in every environment: without CORTEX_DATABASE_URL
the audit stays on stdout and gateway-local tickets are disabled.
License
Available Tools
5 toolslist_cortex_resourcesAInspect
Lists the MCP resources exposed by the gateway (self-describing architecture document, dynamic backend resources through URI templates). Wrapper tool over the resources/list channel, needed because some MCP clients do not surface the resources primitive in their visible tool palette.
Inputs: no named parameters.
list_cortex_ticketsAInspect
Lists your own previously filed report_missing_capability tickets (auto-filtered by OAuth identity — you do not see other agents' tickets). Useful before re-filing a ticket, or to follow a ticket's status after triage by the platform team.
Possible statuses: open (just filed), triaged (the team looked at it), planned (work scheduled), resolved (capability shipped, resolvedPrUrl points at the PR), wont_fix (deliberately not implemented, triageNote explains why), duplicate (merged with another).
Returns: { tickets: [{ id, status, severity, contextTool, contextApp, whatIWanted, suggestedShape, triageNote, resolvedPrUrl, createdAt, updatedAt }] }, sorted by createdAt desc, max 100.
Inputs: limit, status, severity, contextApp.
read_cortex_resourceAInspect
Reads an MCP resource by URI (e.g. cortex://architecture, docs://document/42). Wrapper tool over the resources/read channel. Returns Markdown or JSON depending on the resource. Use after list_cortex_resources to discover available URIs.
Inputs: uri (required).
report_missing_capabilityAInspect
Reports a missing capability or an insufficient tool to the platform team. Use it when you cannot fulfil a user request because no MCP tool offers what is needed, OR when an existing tool limits you (incomplete return, missing parameter, unhelpful error, ...).
Typical example: the user asks for a link to a file; you search the returns of list_files / read_file, there is no webUrl field; you report the missing capability. The platform team prioritizes and ships the field in a next release.
Principle: your ticket is a SIGNAL, not an implementation order. Several converging tickets = a priority. Do not hesitate to report even when unsure — noise is filtered at triage. Do NOT report an obvious bug (the API returns 500) — let the error surface naturally instead.
Auto-dedup: an identical ticket (same description + same tool + same agent) within 24h just refreshes the timestamp instead of creating a duplicate.
Before reporting, consider calling list_cortex_tickets to check you have not already filed a similar ticket — avoids re-filing in a loop.
Returns: { ticketId, deduplicated, ackMessage }.
Inputs: severity (required), userIntent (required), whatIWanted (required), contextApp, +2 more.
whoamiAInspect
Returns the caller's OAuth identity (email, sub, pool, scopes, client, audience, jti) and aggregates the whoami of every healthy federated backend (effective role in that backend, derived capabilities). Useful to know who you are and what you may do before invoking a write tool.
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool has a clearly distinct purpose: listing resources, listing tickets, reading resources, reporting missing capabilities, and identity check. No overlap or ambiguity.
Most tools follow a 'verb_cortex_noun' pattern (list_cortex_resources, list_cortex_tickets, read_cortex_resource), but report_missing_capability lacks 'cortex' and whoami is a single verb, causing minor inconsistency.
5 tools is well-scoped for a gateway providing resource discovery, ticket management, and identity checking; each tool earns its place without bloat or deficiency.
The tool set covers all essential operations: discovering resources, reading them, filing tickets, viewing own tickets, and identity verification. No dead ends or missing functionality for the stated purpose.
Maintenance
Related MCP Connectors
One MCP endpoint for your company's tools, with per-person sign-in, policy and audit.
MCP gateway with runtime security policy, tool-call-level control, and audit of agent actions.
Governed MCP gateway: one endpoint for your tools, with credential custody and audit log.
One OAuth connector for your team's remote MCP servers, with Tool Packs, guardrails and logs.
Related MCP Servers
- FlicenseNot gradedqualityNot gradedmaintenanceAggregates multiple MCP servers behind a single, secure endpoint with unified tool/resource discovery, OAuth authentication, and resilient request routing. Enables users to manage and interact with multiple MCP backends through one centralized interface with load balancing and circuit breakers.2-
- FlicenseNot gradedqualityDmaintenanceA centralized gateway and router that integrates multiple MCP servers into a single endpoint with built-in policy enforcement and secret management. It features a Web GUI for managing tool access, audit logs, and multi-environment configurations across various sub-servers.-
- AlicenseNot gradedqualityCmaintenanceA unified MCP gateway that aggregates multiple MCP servers and API plugins behind a single endpoint with authentication, rate limiting, audit logging, REST API bridge, and web dashboard.8MIT
- AlicenseNot gradedqualityBmaintenanceEnables clients to access multiple backend MCP servers through a single endpoint, with OAuth 2.1 authorization, namespaced tools, and secure credential management.2AGPL 3.0