Serveur MCP MCPG

MCPG — Model Context Protocol Gateway

MCPG is a protocol-authority gateway for the Model Context Protocol. It sits between MCP clients and downstream backend systems, owning protocol handling, session lifecycle, identity, authorization, execution dispatch, and observability.

Production-grade Rust · extensive unit + integration + MCP-conformance coverage · implements MCP 2026-07-28 and 2025-11-25 (both CI-conformance-gated; legacy 2025-06-18 / 2025-03-26 accepted on request)

Build an MCP server with MCPG: see docs/agent-authoring-guide.md for the AI-agent authoring procedure, and docs/mcp-server-ideas.md for 25 concrete systems worth wrapping (name.com, Cloudflare, GitHub, Stripe, iPhone Simulator, and more) — each with a runnable sample under examples/.

Compliance status: start at docs/compliance/mcp-compliance.md for the canonical reference. The supported-feature matrix is at docs/compliance/compliance-support.md; operator knobs and metrics are documented in docs/configuration.md and docs/observability.md. Doc index: docs/README.md.

What MCPG Does

• Speaks MCP over Streamable HTTP with SSE streaming — the finalized 2026-07-28 revision (stateless, header-routed) and the session-based 2025-11-25 default; the server always emits 2026-07-28 as its wire string and accepts the legacy 2025-06-18 / 2025-03-26 revisions when a client requests them
• Routes tool calls, prompt gets, and resource reads to operator-defined backends
• Enforces pre-dispatch authorization with trust levels and CEL expressions
• Resolves identity via JWKS, OIDC/OAuth (multi-provider), or header assertion
• Orchestrates multi-step pipelines with client interaction (elicitation, sampling)
• Supports MCP Tasks — task-augmented tool calls with background execution, polling, cancellation, and result retrieval
• Distributes state across NATS, Redis, or in-process backends per capability (activated via per-capability <capability>.store / <bus>.bus overrides)
• Manages outbound credentials (OAuth 2.0 client_credentials, Vault dynamic DB, static) via the credential_issuer plugin family — backends reference issued values through cred://<plugin_id>/<target> URIs, with cached, auto-refreshed tokens
• Surfaces MCP App URLs (_meta.mcpAppUrl) on resource descriptors with CEL interpolation
• Delivers subject-scoped resource notifications with 4 filter scopes (all, subject, session, expression)
• Extends via async plugin system — dual-tier (Native + Wasm) tool-gate, transform, and identity chains
• Enforces external guardrails via HTTP webhook callouts with CEL triggers and fail-open/fail-closed policies
• Gates tool calls via Machine Payment Protocol with HMAC-bound challenges
• Emits Prometheus metrics and OpenTelemetry traces

Backends

Backends are cdylib plugins — the gateway binary hard-wires none. The catalog ships 32 backend and connector plugins; a caller selects one per capability via a nested backend.kind: discriminator. Common kinds:

Kind	Transport
http	HTTP POST/GET to downstream endpoints
command	Local subprocess execution
nats / kafka / amqp	Messaging (request/reply, pub/sub)
grpc	gRPC via proto-less JSON mapping
graphql / soap	GraphQL query/mutation, SOAP
sql	SQL + warehouse databases (Postgres, MySQL, Snowflake, BigQuery, ClickHouse, DuckDB, Oracle, …)
ldap / dynamodb / email / sftp	Directory, key-value, mail, file transfer
openapi	OpenAPI spec → generated tools
mock	Static fixture response
pipeline	Multi-step orchestration (11 step types)
LLM/media family	17 kinds — OpenAI, Azure OpenAI, Anthropic, Gemini, Stability, and OpenAI-compatible chat/embedding/image/TTS/STT

Full reference: docs/backends.md.

Install

The model-context-protocol-gateway package installs mcpg together with mcpg-config, mcpg-cloud and mcpg-plugin:

npx model-context-protocol-gateway --help       # npm, one-shot
npm install -g model-context-protocol-gateway   # npm
pipx install model-context-protocol-gateway     # PyPI (or: uvx model-context-protocol-gateway)
cargo install model-context-protocol-gateway --version <version>    # crates.io: a launcher that runs the release binaries
cargo binstall model-context-protocol-gateway --version <version>   # the release binaries directly
curl -fsSL https://raw.githubusercontent.com/mcpg-dev/mcpg/main/install.sh | sh

The container image is ghcr.io/mcpg-dev/mcpg:<version>. Every install path: mcpg.dev/docs/install.

Build from source

# Build
cargo build

# Test
cargo test

# Run
MCPG_CONFIG=config.example.yaml cargo run

Docker & Kubernetes

# Build Docker image (from repo root)
docker build -t mcpg:latest -f apps/gateway/Dockerfile .

# Run single instance
docker run -p 8787:8787 -v $(pwd)/config.yaml:/etc/mcpg/config.yaml:ro mcpg:latest

# Deploy to Kubernetes via Helm (single instance). The charts are
# 0.1.0-beta.N prereleases: always pass --version.
helm install mcpg oci://ghcr.io/mcpg-dev/charts/mcpg --version <chart-version>

# Deploy HA with NATS backend (multi-instance)
helm install mcpg oci://ghcr.io/mcpg-dev/charts/mcpg --version <chart-version> \
  --set replicaCount=3 \
  --set nats.enabled=true \
  --set autoscaling.enabled=true \
  --set podDisruptionBudget.enabled=true

Full deployment guide: docs/deployment.md. Helm chart reference: Kubernetes install with Helm.

Configuration

See config.example.yaml for a working configuration. Full reference: docs/configuration.md.

server:
  bind_address: "127.0.0.1:8787"
observability:
  logs:
    level: "info"
    sinks:
      - kind: stderr
        config: { format: json }
mcp:
  capabilities:
    tools:
      - name: "my_tool"
        description: "Example HTTP backend"
        backend:
          kind: http
          url: "https://api.example.com/endpoint"
          method: post
          timeout_ms: 5000

Documentation

Document	Description
docs/architecture.md	System architecture, module structure, request flow
docs/request-flow.md	Complete request lifecycle, all routes, plugin extension points
docs/configuration.md	Complete configuration reference
docs/backends.md	Backend kinds and connector plugins with examples
docs/pipelines.md	Pipeline execution, 11 step types, suspension/resumption
docs/identity-and-authorization.md	JWKS, OIDC/OAuth, trust levels, CEL policy
docs/infrastructure.md	Session stores, delivery buses, pipeline stores
docs/observability.md	Logging, Prometheus metrics, OpenTelemetry
docs/api-reference.md	HTTP endpoints, MCP operations, SSE protocol
docs/deployment.md	Docker, Kubernetes/Helm, deployment topologies
docs/plugins.md	Plugin system: traits, registry, Wasm, config
Kubernetes install with Helm	Helm chart reference — parameters, examples, architecture

Architecture at a Glance

MCP Client
    │
    ▼
HTTP Transport ──→ Identity Resolution ──→ Gateway Runtime ──→ Plugin Chain ──→ Backend Execution
    │                (OIDC/JWKS/Plugin)           │           (gate/transform)   (backend plugins)
    │                                            │
    ├── SSE Streaming                    Session Management
    │   (replay window)                  (4 store backends)
    │
    └── Prometheus + OpenTelemetry

Plugin System Crates

libs/
├── plugin-protocol/  — Trait contracts + FFI ABI types (ToolGatePlugin, BackendPlugin, etc.)
├── plugin-host/      — Registry, chain evaluation, Wasm hosting, artifact verification
├── plugin-sdk/       — SDK for building plugins (testing helpers, MockGateway)
├── sdk/              — `mcpg-sdk` umbrella crate plugin authors depend on
└── plugins/          — the production catalog: 92 shipping plugins across 14
                        categories (backend, security, identity, transform,
                        observability, payment, credential, …), cdylib + a few Wasm

Design Decisions

Key architectural decisions:

• ADR-0001: Single-deployment, backend-driven runtime (no multi-tenancy)
• ADR-0004: Canonical backend execution contract (gateway as protocol authority)
• ADR-0005: Multi-node lease fencing with CAS semantics
• ADR-0007: OIDC/OAuth multi-provider authentication

License

The gateway core is Apache-2.0 (see LICENSE). Enterprise plugins and the fleet platform (Kubernetes operator, control plane) ship under BUSL-1.1, each release auto-converting to Apache-2.0 three years later. To license production use of the BUSL-1.1 components, email agent@mcpg.dev or use the managed service at mcpg.cloud. The canonical component map is at mcpg.dev/license.

Backend governance integrates with the policy layer. Each backend's minimum_trust and allow_if are injected as per-tool policy rules at startup. CEL expressions are compiled and validated during bootstrap.

Debug Tools

The debug capability model is intentionally narrow:

• the top-level debug.enabled: true gate must be set before any debug capabilities exist
• command probes, network probes, and the debug HTTP JSON call use named execution profiles (under debug.tools.command_profiles / debug.tools.network_profiles) with bounded timeout and size limits
• built-in debug tools, prompt, and resource capabilities can be exposed or hidden independently via debug.tools.exposure.*
• debug tools emit structured logs with backend_kind: debug_tool to distinguish from operator backends

Execution Contract

All downstream execution paths produce structured result envelopes with:

• tool name, profile, and request kind classification
• request and response details
• structured error objects with category, retry hints, idempotency, and suggested actions
• execution duration, status codes, and truncation indicators

Execution logs include backend_kind (operator_backend or debug_tool) on all call sites.