mcp-hub

Gateway + registry for a fleet of MCP servers. One endpoint, many servers, per-tenant RBAC, audit, rate limits. Stateless per the 2026-07-28 spec.

┌──────────┐   POST /   ┌─────────┐   tools/list   ┌────────────────┐
│ Claude   │──────────▶ │ mcp-hub │ ─────────────▶ │ github-mcp     │
│ Cursor   │──────────▶ │ (router)│   tools/call   │ jira-mcp       │
│ any MCP  │──────────▶ │         │ ─────────────▶ │ db-tools       │
│ client   │  Bearer    └─────────┘                │ ... 10 servers │
└──────────┘  + RBAC                               └────────────────┘

The problem this kills: agents in production hold connections to 4-10 MCP servers with separate auth, latencies and failure modes. Tool discovery becomes the bottleneck — 60+ tools flood the context window, one tools/list round-trip per server per refresh. MCP broke 97M monthly SDK downloads with no production playbook; this is the playbook's first chapter.

Install & run

npm install -g @threadwire/mcp-hub

mcp-hub init                       # ~/.mcp-hub/config.json (binds 127.0.0.1 only)
mcp-hub add github https://mcp.github.com --tools issues.create,pr.list --scopes read,write
mcp-hub add jira   https://mcp.atlassian.dev --tools issue.transition
mcp-hub doctor                     # health-check every upstream
mcp-hub start                      # http://127.0.0.1:8801

Point any MCP client at http://127.0.0.1:8801 with authorization: Bearer <tenant-token>.

curl -s http://127.0.0.1:8801 \
  -H 'authorization: Bearer mcp-hub-dev-token' \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"issues.create","arguments":{"title":"hello"}}}'

What's inside

| Module | Job | |-------------------|---------------------------------------------------------------------| | router | stateless JSON-RPC endpoint; no handshake, no session pinning | | registry | server + tool catalog, aggregated tools/list with TTL cache | | proxy | pure JSON-RPC 2.0 to upstream Streamable HTTP; requestState per call | | auth | Bearer → tenant; RFC 8707 resource check; RFC 9207 iss validation | | rbac | per-tenant allow/deny globs — least privilege, deny wins | | rate | sliding window per tenant; retryAfterMs on 429 | | audit | JSONL audit: who/what/input-hash/status/latency — raw args never logged | | cache | ttlMs + cacheScope discovery cache per the 2026-07-28 spec | | discovery | boot-time syncAll() — upstream tool-list pull, non-fatal on error | | store | SQLite config/tenants/audit via built-in node:sqlite | | circuit | per-upstream circuit breaker (open / half-open / closed) | | oauth / plugin | OAuth PKCE upstream auth · schema transformer hooks |

Why it earns trust

• Stateless — round-robin behind any load balancer, serverless-friendly, no Redis session store.
• Input-hashed audit — OWASP's top MCP risk category ("Lack of Audit and Telemetry") closed by default; sensitive args are SHA-256 fingerprints, never payloads.
• Never binds 0.0.0.0 — the June 2025 mass-exposure lesson is baked into the default config.
• Structured errors with dedup keys — a timed-out call can be retried with requestState; the gateway never lets an agent believe a tool ran when it didn't.
• Zero runtime deps — pure Node, one compiled binary worth of code.
• Distributed traces built-in — W3C traceparent is relayed verbatim through the gateway to every upstream; pair with mcp-telemetry for end-to-end spans.

Observability

Point the hub at mcp-trace and every call lands in one live dashboard:

{ "telemetryUrl": "http://127.0.0.1:8901", "telemetryToken": "trace-hub-secret" }

mcp-trace --serve 8901 --token trace-hub-secret &   # from mcp-telemetry
mcp-hub start                   # dashboard: http://127.0.0.1:8801/admin

telemetryToken is optional but recommended — the feed's --serve rejects /ingest without a matching --token (or MCP_TRACE_TOKEN), and the hub sends it as Authorization: Bearer ... on each span POST. When unset, spans still flow to a feed running without a token.

The hub posts one span per tools/call outcome — OK, DENIED, RATE_LIMITED, circuit-open, or upstream error — whose traceId and parent_span_id follow the inbound W3C traceparent, so client → hub → upstream chains link end to end in the dashboard. The bridge is fire-and-forget: a down telemetry feed never slows a call. Audit rows are pruned after auditRetentionMs (default 30 days, 0 disables), and stale tools/list caches can be flushed with POST /cache/invalidate (or automatically on POST /admin/sync).

Docker

One command brings up the pair — gateway + telemetry, wired together:

docker compose up --build     # hub :8801 · dashboard :8901

• mcp-hub/deploy/hub-config.json is the config you edit (mounts read-write so mcp-hub add still works).
• Default MCP_TRACE_TOKEN=trace-hub-secret — override with an env var, and keep hub-config.json's telemetryToken in sync.
• Both images run unprivileged; telemetry persists its feed to a volume at /data; hub's registry/audit lives at /data too.

Use as a library

The gateway is plain classes — embed it without the CLI:

import { HubServer, Router } from "@threadwire/mcp-hub";

// transport-agnostic core: attach handle() to any Node HTTP stack
const router = new Router(cfg, "/tmp/mcp-hub.db");
server.on("request", (req, res) => {
  backend(await router.handle(JSON.parse(body), req.headers), res);
});

// or the whole gateway, HTTP + admin + oauth included
await new HubServer(cfg, "/tmp/mcp-hub.db").listen(8801, "127.0.0.1");

Public surface: HubServer, Router, HubStore, Discovery, Pipeline, runStdioBridge, HUB_VERSION, PROTOCOL_VERSION, plus all config/types. Router owns zero I/O except its SQLite store and fetch calls — exact same code paths the CLI runs. Full runnable example: examples/embed.mjs.

npm test    # unit + integration against an in-repo fixture upstream
mcp-hub audit --n 50   # tail the audit

MIT.

I hope everyone will help develop MCP-HUB by submitting pull requests.