Open-source MCP gateways, compared
Eight self-hosted gateways that sit between your agents and their MCP servers: what each one controls, who it suits, and what to check before you run it.
Checked against each project's repository on .
An MCP gateway is the single endpoint your agents connect to instead of every MCP server directly. It decides which servers and tools a caller can see, holds the credentials those servers need, and records what was called. That one choke point is where most MCP governance gets enforced.
An open-source gateway gives you that control point without a vendor in the request path: you can read the policy code, run it inside your own network and keep every log. In exchange, you own patching, scaling and failover.
The eight projects below are the open-source gateways most often evaluated for MCP. They differ in what they are for, from a developer's laptop to a governed, multi-tenant platform, so the list says who each one suits rather than ranking them. Licenses, languages and features are as each project's repository and documentation state them; check the linked repository for the current release before you decide.
At a glance
| Gateway | Best for | License | Language |
|---|---|---|---|
| agentgateway | Platform teams that want one policy-driven data plane for MCP, agent-to-agent and LLM traffic. | Apache-2.0 | Rust |
| IBM ContextForge | Teams that need to turn existing REST and gRPC services into curated MCP tools. | Apache-2.0 | Python |
| Obot | Organizations that want a full MCP governance platform: registry, hosting, access by group and audit. | MIT | Go |
| Microsoft MCP Gateway | Azure and Entra ID shops running their MCP servers on Kubernetes. | MIT | C# |
| Docker MCP Gateway | Standardizing the MCP servers developers run on their own machines. | MIT | Go |
| MCPJungle | Small platform teams that want a simple self-hosted registry and gateway with per-client allow-lists. | MPL-2.0 | Go |
| Unla | Exposing existing internal HTTP APIs as MCP tools without writing an MCP server. | MIT | Go, with a web management UI |
| Lasso MCP Gateway | Security teams adding data masking and pre-load server scanning in front of local MCP servers. | MIT | Python |
The gateways
01 · Linux Foundation project
agentgateway
Best for: Platform teams that want one policy-driven data plane for MCP, agent-to-agent and LLM traffic.
- License
- Apache-2.0
- Language
- Rust
- Runs on
- Standalone with a YAML config, or on Kubernetes with its built-in controller and the Gateway API
A proxy built for agent protocols. As an MCP gateway it federates tools from many servers behind one endpoint, speaks stdio, HTTP, SSE and Streamable HTTP, and can expose OpenAPI services as MCP tools. The same proxy also routes A2A (agent-to-agent) traffic and LLM calls.
Governance controls it documents
- Authentication with JWT, API keys or OAuth
- Fine-grained RBAC written as CEL policies
- Rate limiting and TLS
- Guardrails: regex, OpenAI moderation, AWS Bedrock Guardrails, Google Model Armor or your own webhook
- OpenTelemetry metrics, logs and traces
Check before adopting
- The project describes itself as in active development: pin a release and read its changelog before upgrading.
- It covers far more than MCP, so scope the first rollout to the MCP routes you need.
02 · IBM
IBM ContextForge
Best for: Teams that need to turn existing REST and gRPC services into curated MCP tools.
- License
- Apache-2.0
- Language
- Python
- Runs on
- PyPI package or container; multi-cluster on Kubernetes with Redis-backed federation and caching
A registry and proxy that federates MCP servers, A2A agents and REST or gRPC APIs into one endpoint. It translates gRPC services to MCP by reflection, wraps legacy APIs as MCP tools, and lets you bundle chosen tools into virtual servers so each client sees only the set made for it.
Governance controls it documents
- Authentication with Basic, JWT or custom schemes, plus user-scoped OAuth tokens
- Rate limiting and retries per upstream
- Virtual servers that expose a curated tool list
- More than 40 plugins for extra transports, protocols and integrations
- OpenTelemetry tracing to Phoenix, Jaeger, Zipkin or any OTLP backend
- Admin UI, with support for air-gapped deployments
Check before adopting
- It will not start without JWT_SECRET_KEY and AUTH_ENCRYPTION_SECRET set, in every environment: plan where those secrets live.
- It is a Python service; measure the latency it adds under your own load before putting it in front of latency-sensitive agents.
03 · Obot AI
Obot
Best for: Organizations that want a full MCP governance platform: registry, hosting, access by group and audit.
- License
- MIT
- Language
- Go
- Runs on
- Kubernetes for production; a Docker setup for evaluation and trusted single-tenant use
More than a proxy: Obot combines an MCP gateway with an MCP and skills registry built on Git-backed catalogs, hosting for npx, uvx and containerized MCP servers, an LLM gateway, and audit logs. Agents reach servers through virtual MCP servers that combine catalog entries behind one endpoint.
Governance controls it documents
- Server and tool access by user or identity-provider group, through virtual MCP server profiles
- Each component pinned to a catalog snapshot, so changes arrive only through an explicit upgrade
- Inspect, reject or modify MCP requests and responses with MCP, webhook or built-in credential filters
- MCP tunnels to reach servers on private networks without opening an inbound port
- User provisioning from Okta through SCIM
- Correlated audit logs, exportable once or on a schedule
Check before adopting
- Community and Enterprise editions run the same open-source code and differ by license key: confirm which of the controls you need are in the edition you plan to run.
- The Docker configuration mounts the host Docker socket; the project says to use Kubernetes for production.
04 · Microsoft
Microsoft MCP Gateway
Best for: Azure and Entra ID shops running their MCP servers on Kubernetes.
- License
- MIT
- Language
- C#
- Runs on
- Kubernetes, locally or on Azure through a one-click Bicep deployment
A reverse proxy and management layer for MCP servers in Kubernetes. Its control plane is a REST API that deploys, updates and deletes MCP servers (adapters) and registered tools; its data plane routes each MCP request on its own to a ready server instance, with a tool router that sends tool calls to the server that owns them.
Governance controls it documents
- Entra ID authentication on both the data plane and the control plane
- Authorization by Entra app roles: read for a resource's creator and the roles it names, write for its creator and mcp.admin
- Server lifecycle (deploy, update, delete) through one API instead of ad-hoc deployments
- Telemetry and observability integration points
Check before adopting
- The current version requires clients and adapters on the MCP 2026-07-28 protocol, with no fallback for older clients: read its migration guide first.
- Its identity model is Entra ID; teams on another identity provider should check what that would take.
05 · Docker
Docker MCP Gateway
Best for: Standardizing the MCP servers developers run on their own machines.
- License
- MIT
- Language
- Go
- Runs on
- The docker mcp CLI plugin, inside Docker Desktop's MCP Toolkit or on its own with Docker CE
The engine behind Docker Desktop's MCP Toolkit. Each MCP server runs in its own container, and every client (VS Code, Cursor, Claude Desktop and others) connects to the same gateway configuration. Servers are grouped into profiles that can be shared through OCI registries.
Governance controls it documents
- Container isolation for every MCP server, with minimal host privileges for npx and uvx servers
- Secrets kept in Docker Desktop's secrets management rather than environment variables
- Built-in OAuth flows for servers that need them
- Catalogs and shareable profiles that define which servers a team uses
- Logging and call tracing
Check before adopting
- It is built around Docker Desktop; some features, such as profiles, must be switched on when it runs without it.
- If you need a shared, central gateway with per-user authorization, check its documentation for that model before choosing it over a server-side gateway.
06 · MCPJungle project
MCPJungle
Best for: Small platform teams that want a simple self-hosted registry and gateway with per-client allow-lists.
- License
- MPL-2.0
- Language
- Go
- Runs on
- Self-hosted; Docker Compose for a quick start, a CLI to register servers
Register each MCP server once and point Claude, Cursor, Codex or your own agents at a single endpoint. It runs in a development mode for personal use and an enterprise mode for shared deployments, where the governance controls switch on.
Governance controls it documents
- In enterprise mode, no client reaches any server until it has an identity and an allow-list of servers
- Tool groups that expose only the tools a client should see
- Prometheus-compatible metrics through OpenTelemetry, on by default in enterprise mode
Check before adopting
- Development mode lets every connected client reach every registered server: make sure shared deployments run in enterprise mode.
- MPL-2.0 is a file-level copyleft license: have counsel confirm what it means if you modify and distribute it.
07 · AmoyLab
Unla
Best for: Exposing existing internal HTTP APIs as MCP tools without writing an MCP server.
- License
- MIT
- Language
- Go, with a web management UI
- Runs on
- Docker, bare metal, VMs or Kubernetes with Helm
A lightweight gateway that turns existing APIs into MCP servers through YAML configuration, with no code changes, and also proxies existing MCP servers. It serves SSE and Streamable HTTP endpoints and ships a management UI.
Governance controls it documents
- Multi-tenant support
- OAuth-based pre-authentication for MCP servers
- Multiple replicas, with configuration synced over OS signals, HTTP or Redis Pub/Sub
Check before adopting
- The project says it is under rapid development and cannot guarantee backward compatibility: read the release notes before each upgrade.
- Every API you expose becomes a tool an agent can call: review each mapping the way you would a new API client.
08 · Lasso Security
Lasso MCP Gateway
Best for: Security teams adding data masking and pre-load server scanning in front of local MCP servers.
- License
- MIT
- Language
- Python
- Runs on
- Installed with pip or built as a container; launched by the MCP client from its mcp.json
A plugin-based gateway that wraps the MCP servers in a client's configuration, manages their lifecycle and inspects what passes through. Its security scanner checks a server's reputation and tool descriptions before the server is loaded.
Governance controls it documents
- A basic guardrail plugin that masks tokens in requests and responses
- PII detection and anonymization through Microsoft Presidio
- A security scanner for server reputation and tool descriptions before loading
- Tracing through the xetrack plugin
Check before adopting
- It runs alongside each client rather than as one shared network service, so central policy means managing every client's configuration.
- The optional Lasso guardrail sends content to Lasso's API for checking: review that data flow before enabling it.
- Check the repository's recent commits and releases for how actively it is maintained.
How to choose
Who is calling, and as whom?
A gateway is only as good as the identity it enforces. Check whether it authenticates each agent and user against your identity provider, or lets everything share one key, and whether a tool call can be traced back to the person it acted for. Read more: Agent Identity and Role Binding as the Foundation of MCP Tool Access Control.
Can policy say no before the call runs?
Look for authorization that can tell callers, servers and individual tools apart, and that filters what a client discovers, not just what it can call. A tool an agent never sees is a tool it cannot be tricked into using. Read more: Centralizing MCP Access Policy Enforcement Without Fragmenting It Across Individual Servers.
Where does it run, and who runs it?
A desktop wrapper, a Kubernetes data plane and a full platform carry very different operating costs. Open source means your team owns upgrades, scaling and failover; size that work before you size the license. Read more: MCP Gateway Architecture Patterns for Enterprise Scale.
What will you be able to prove afterwards?
Auditors ask what an agent did, not what it was allowed to do. Check that the gateway records who called which tool with what result, and exports it to the logging and tracing stack you already run. Read more: Closing the Observability Gap Between What MCP Agents Request and What They Actually Retrieve.
Does it inspect what comes back?
Tool poisoning and prompt injection arrive in tool descriptions and responses. Guardrails, response filters and server vetting at the gateway are cheaper than cleaning up after an agent acts on bad input. Read more: Tool Poisoning Attack Patterns in Production MCP Deployments.
Frequently asked questions
What is an MCP gateway?
A proxy between AI agents and the MCP servers they use. Agents connect to the gateway instead of to each server, and the gateway handles discovery, authentication, authorization, credentials and logging for every tool call in one place.
Is an open-source MCP gateway enough for enterprise governance?
It can carry the controls (identity, authorization, filtering and audit), but governance also needs an inventory of approved servers, an owner for each and a process for onboarding new ones. A gateway enforces those decisions; it does not make them.
Do I need a gateway if my MCP servers already use OAuth?
OAuth on each server proves who is calling that server. A gateway adds one place to decide which servers and tools each caller may reach, to keep credentials out of clients and to log every call the same way, instead of repeating that work in every server.
Which of these gateways is fastest?
It depends on your workload more than on any published number, and the language a gateway is written in is only part of it. Run each candidate in front of your own servers, at your own concurrency, and measure the latency it adds.
What does an open-source MCP gateway cost?
The software is free under the license listed for each project. The cost is the infrastructure it runs on and the people who operate it; some projects also sell an enterprise edition or support, priced on their own sites.
Go deeper
- MCP Gateway Evaluation Criteria for Enterprise Buyers
Five criteria separate enterprise MCP gateways from proxies that collapse under compliance pressure.
- How to Evaluate an AI Agent Gateway for Enterprise Use
Distinguish runtime blocking from monitoring-only when evaluating AI gateways.
- MCP Gateway Controls Every Enterprise Deployment Should Have
Security controls that prevent agents from exploiting unprotected MCP servers.
- MCP Gateway Architecture Patterns for Enterprise Scale
Stateless protocol redesign enables horizontal scaling without sticky sessions or shared state.
- Centralizing MCP Access Policy Enforcement Without Fragmenting It Across Individual Servers
A centralized gateway enforces consistent MCP access policy across all servers and teams.
- Hosting MCP Servers on Kubernetes in the Enterprise
Kubernetes provides the governance and resilience layer enterprise MCP deployments urgently need.
- Evaluating Purpose-Built MCP Governance Platforms Against Patchwork Privacy Tooling for Agentic Workflows
Purpose-built governance platforms enforce MCP access controls where agents act.