MCP Conduit

The open-source MCP gateway that makes AI tool calls production-ready.

Conduit sits between your AI agents and MCP servers. One YAML file gives you caching, auth, rate limiting, guardrails, and multi-server routing. No code changes needed.

  AI Agents ──> [ Conduit ] ──> MCP Server A (HTTP)
  (Claude,                 ──> MCP Server B (stdio)
   Cursor,                 ──> MCP Server C (replicas)
   custom)

Get Started in 60 Seconds

npx conduit init   # interactive setup wizard
npx conduit        # start the gateway

Or create conduit.config.yml:

servers:
  - id: my-server
    url: http://localhost:3000/mcp
    cache:
      default_ttl: 300

npm install && npm run build && npm start

Your agents now connect to http://localhost:8080/mcp/my-server instead of the MCP server directly. Open http://localhost:8080/conduit/dashboard to see what's happening.

Performance

Measured on a single machine, single process. These are real numbers, not theoretical limits.

| What we measured | Result | Why it matters | |---|---|---| | Gateway overhead | 0.02ms added per request | You won't notice Conduit is there | | Cache hit throughput | 46K RPS (sequential) / 64K RPS (concurrent) | Cached responses are instant | | All features enabled | 111K RPS at P99 = 3.24ms | Auth + plugins + rate limit + cache, still fast | | Stdio transport | 354K RPS (concurrent) | Faster than HTTP (no network overhead) | | Plugin hooks | 2M ops/sec | Plugins add zero measurable latency |

| | Conduit | Bifrost (Go) | TrueFoundry | Envoy AI GW | |---|---|---|---|---| | Overhead | 0.02ms | 0.011ms | 3-4ms | 1-2ms | | Cache hit RPS | 46K-64K | not published | 350/vCPU | not published | | Language | TypeScript | Go | TypeScript | C++/Go | | License | MIT | Apache 2.0 | Proprietary | Apache 2.0 |

Bifrost has lower raw overhead (Go vs JS). We compensate with aggressive caching. In practice, most requests are cache hits.

npm run benchmark         # run it yourself (24 scenarios)
npm run benchmark:quick   # 3-second smoke test

What Conduit Does

Core Gateway

| Feature | Description | |---|---| | Multi-server routing | Route tools across servers with namespace prefixes (github.list_repos) | | HTTP + stdio transport | Proxy HTTP servers or spawn local binaries. Same features for both | | Load balancing | Round-robin or least-connections across replicas | | Circuit breaker | Per-replica failure isolation (closed/open/half-open) | | Health checks | Periodic probes with configurable thresholds | | Hot reload | SIGHUP or API call. Add servers, update rules, zero downtime |

Caching

| Feature | Description | |---|---| | L1 in-memory | LRU cache with per-tool TTL | | L2 Redis | Distributed cache for multi-instance deployments | | Smart invalidation | Destructive tools auto-invalidate related cache entries | | Inflight dedup | 10 identical requests = 1 upstream call, 10 responses | | Annotation-aware | Respects MCP readOnly, idempotent, destructive hints |

Security

| Feature | Description | |---|---| | JWT auth | JWKS validation with auto-rotating keys | | API key auth | Constant-time comparison (timing-attack resistant) | | ACL policies | Per-client, per-server, per-tool access control with wildcards | | AI guardrails | Block/alert/transform tool calls based on name + argument rules | | Rate limiting | Sliding window (memory or Redis), global + per-client + per-tool | | CSRF protection | Admin mutations require X-Conduit-Admin header | | SSRF protection | URL validation blocks internal IPs and metadata endpoints |

Observability

| Feature | Description | |---|---| | Structured logs | SQLite with retention, field redaction, trace correlation | | Prometheus metrics | 18 counters/gauges/histograms on dedicated port | | OpenTelemetry | W3C Trace Context + OTLP export | | Admin dashboard | React SPA: logs, servers, cache stats, circuit breakers | | Admin API | 40+ endpoints for programmatic access |

Extensibility

| Feature | Description | |---|---| | Plugin system | 5 hooks in the pipeline, dynamic JS/TS loading | | Service discovery | HTTP self-registration + DNS SRV | | JSON Schema | Config autocompletion in VS Code / JetBrains | | CLI wizard | npx conduit init generates config interactively |

Configuration

Minimal (just your server):

servers:
  - id: my-server
    url: http://localhost:3000/mcp
    cache:
      default_ttl: 300

With stdio transport:

servers:
  - id: filesystem
    transport: stdio
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    cache:
      default_ttl: 60

With everything:

servers:
  - id: salesforce
    url: http://salesforce-mcp:3000/mcp
    cache:
      default_ttl: 300
      overrides:
        create_lead:
          ttl: 0
          invalidates: [get_contact, search_leads]

auth:
  method: api_key
  api_keys:
    - key: ${CONDUIT_API_KEY}
      client_id: my-agent
      tenant_id: default

guardrails:
  enabled: true
  default_action: allow
  rules:
    - name: block-destructive
      tools: ["delete_*", "drop_*"]
      action: block
      message: "Destructive tools require manual review"
      severity: high

cache:
  l2:
    enabled: true
    redis_url: ${CONDUIT_REDIS_URL}
    default_ttl_multiplier: 3

admin:
  key: ${CONDUIT_ADMIN_KEY}

All config fields: docs/CONFIGURATION.md | Environment variables: .env.example

Deployment

Docker:

docker build -t conduit .
docker run -d -p 8080:8080 -p 9090:9090 \
  -v ./conduit.config.yml:/app/conduit.config.yml:ro \
  -e CONDUIT_ADMIN_KEY=$(openssl rand -hex 32) \
  conduit

Kubernetes: Helm chart in deploy/helm/. Liveness and readiness probes included.

helm install conduit deploy/helm/conduit --set config.adminKey=$(openssl rand -hex 32)

What Conduit Does NOT Do

We believe in being upfront:

• Not a managed service. You host it. No SaaS offering.
• Not a cluster. Single process. Scale horizontally with a load balancer.
• No pre-built integrations. This is a proxy, not an integration catalog.
• No container isolation for stdio. Child processes run in the same OS context.
• No SOC 2 certification. Regulated environments need your own audit.
• Higher raw overhead than Go. 0.02ms vs 0.011ms (language gap). Caching makes this irrelevant in practice.

Documentation

| Doc | Description | |---|---| | Configuration Reference | Every config field, with defaults and examples | | API Reference | All 40+ admin endpoints with curl examples | | Security Guide | Production hardening checklist | | Wiki | Architecture, plugin dev, FAQ | | JSON Schema | Config autocompletion for your editor |

Testing

npm test                    # 1341 tests
npm run test:battle         # chaos + fault injection
npm run test:integration    # multi-server end-to-end

70 test files covering unit, e2e, battle, integration, security, and benchmarks.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

License

MIT