Best MCP Gateways 2026: Open-Source Options Compared
The short answer
Choose an MCP gateway by where it runs and how much governance you need. All five below are open source; versions and stars were read from GitHub on November 19, 2026.
| Gateway | Licence · version | Stars | Runs on | Speaks | Auth and policy | Best for |
|---|---|---|---|---|---|---|
| IBM ContextForge | Apache-2.0 · v1.1.1 (Nov 10, 2026) | ~4,900 | PyPI, Docker, Kubernetes (Helm), Redis federation | MCP, A2A, REST and gRPC → MCP | SSO, RBAC, guardrails, OpenTelemetry | A company-wide tool registry |
| agentgateway | Apache-2.0 · v1.7.0 (Oct 27, 2026) | ~6,300 | Standalone binary or Kubernetes (Gateway API) | MCP, A2A, OpenAPI → MCP, LLM traffic | JWT, API keys, OAuth, CEL-based RBAC, rate limits | Platform teams on Kubernetes |
| Docker MCP Gateway | MIT · no tagged release (last commit Oct 26, 2026) | ~1,600 | Docker Desktop (MCP Toolkit) or docker mcp CLI | MCP | Built-in OAuth flows, per-server containers | Developers running local servers safely |
| Microsoft MCP Gateway | MIT · v1.0.1 (Nov 16, 2026) | ~900 | Kubernetes (StatefulSets), Azure | MCP | Bearer token, RBAC on data and control plane | Azure and AKS shops |
| Bifrost | Apache-2.0 · Helm chart v2.1.49 (Nov 19, 2026) | ~9,800 | Single Go binary, Docker, Kubernetes | OpenAI-compatible LLM API plus MCP | Virtual keys, budgets, OIDC on enterprise | One gateway for models and tools |
1. IBM ContextForge — best central registry
ContextForge is a registry and proxy that federates MCP servers, A2A agents and REST or gRPC APIs behind one endpoint. Its distinctive feature is virtualisation: point it at an existing REST API and it exposes it to agents as MCP tools, so you don’t write a server per internal service. It runs from PyPI or Docker for a pilot and scales to multi-cluster Kubernetes with Redis-backed federation. Pick it when the problem is “hundreds of tools, many teams, one catalogue”.
2. agentgateway — best for Kubernetes platform teams
agentgateway is a proxy built for agent protocols: MCP tool federation over stdio, HTTP, SSE and Streamable HTTP; A2A between agents; and LLM routing, including Kubernetes Inference Gateway extensions that route by GPU utilisation and KV cache. Policy is written in CEL, so per-tool rules (“only the billing agent may call refund”) are code-reviewed like everything else. It runs standalone from a YAML file or as a controller with the Gateway API.
3. Docker MCP Gateway — best on a developer machine
Docker’s gateway runs every MCP server from the Docker MCP Catalog in its own container, with minimal host privileges for npx and uvx servers, and handles OAuth for servers that need tokens. It ships inside Docker Desktop’s MCP Toolkit or as the docker mcp CLI plugin. It is the simplest way to stop MCP servers from running with your laptop’s full permissions, but it is not a multi-team governance layer.
4. Microsoft MCP Gateway — best for Azure
A Kubernetes-native reverse proxy with a control-plane REST API to deploy and manage MCP servers, stateless routing, and RBAC on both planes; metadata sits in Redis locally or Cosmos DB in the cloud. It reached v1.0.1 on November 16, 2026. Choose it if your agents already run on AKS and your identity is Entra ID.
5. Bifrost — best when one gateway handles models and tools
Bifrost started as a high-throughput LLM gateway (one OpenAI-compatible API in front of 23+ providers, failover, semantic caching, budgets) and added MCP gateway features, including “Code Mode”. If you already need a model gateway, running tools through the same layer gives one place for keys, budgets and logs. Some governance features (clustering, OIDC provisioning) sit in the enterprise edition.
Avoid for new projects: MetaMCP (metatool-ai/metamcp) is archived on GitHub as of November 2026.
How to choose in four questions
- Laptop or shared service? Laptop → Docker MCP Gateway. Shared → the rest.
- Do you need A2A or REST-to-MCP? Yes → ContextForge or agentgateway.
- Where does it run? Kubernetes → agentgateway or Microsoft; Azure → Microsoft; anywhere with Python → ContextForge.
- Do you also route LLM calls? Yes → Bifrost or agentgateway, to keep one gateway.
Whichever you pick, keep tool credentials in the gateway rather than in each client config; see best auth platforms for AI agents and MCP servers. If you are deciding between a self-hosted gateway and a managed one, see managed vs self-hosted vs hybrid agent gateway vendors.
Last verified: November 19, 2026.