11 Best Open-Source MCP Gateways in 2026

Self-hosting an MCP gateway is easy. Running one for a whole company is not.

Most open-source gateways do the first job well: put several MCP servers behind one endpoint on your own machine or cluster. Fewer handle what a company needs next, which is people signing in with SSO, each tool call running as the person behind it, and a log you can hand to security.

This guide is for engineering and platform teams choosing an MCP gateway they can run themselves. Here, open source means the code is public on GitHub and you can self-host it. It compares eleven projects and says which fits which team.

What makes an MCP gateway ready for a company, not just a laptop

• SSO and SCIM. People sign in with your identity provider, and access goes away when they leave.
• Per-user OAuth. Tool calls run with each person's own connection, not one shared token for the whole gateway.
• Tool-level access. Admins choose which tools each group can call, including read-only sets.
• An audit log per person. Every call records who made it, through which agent, and what happened.
• Remote and local servers. Vendor-hosted remote servers and local stdio servers sit behind the same rules.
• An upgrade path. You can start with one team and grow without swapping the gateway.

Architecture: local proxy vs remote gateway vs managed platform

A local proxy runs next to one AI client. Docker MCP Gateway and MCPJungle in local mode work this way. Setup is quick, and nothing is shared between people.

A remote gateway is one endpoint the whole team uses. IBM ContextForge, agentgateway, Kuadrant and Microsoft MCP Gateway are built for this, and you connect identity, policy and log storage yourself.

A managed platform includes identity, per-person access and audit logging, whether you self-host it or use the project's cloud. Metorial, Obot and Archestra sit here.

Problems the right gateway prevents

Shadow MCP. Engineers install servers on their laptops to get work done, and nobody knows which tools reach company data. A shared gateway gives them an approved path. More in what is shadow AI.

API keys in config files. A personal token in a local config works, gets copied to teammates, and outlives the person who created it. Per-user OAuth removes the reason to paste it.

No audit trail. When an agent changes the wrong record, the question is whose access it used. Without a per-person log, nobody can answer. See are MCP servers secure.

One shared credential for everyone. Several gateways store a single OAuth token for the whole gateway. Every agent then reaches everything that token can, whoever is asking.

How to evaluate an open-source MCP gateway: 5 criteria

Identity: SSO, SCIM and per-user OAuth

Check whether SSO is in the open-source version or only in a paid edition, and whether OAuth tokens are stored per person or once for the gateway.

Access control: per-tool allow and deny, read-only modes

Look for rules by group or role down to single tools. Some gateways only let you enable or disable a tool for everyone.

Visibility: an audit log per person, SIEM export

Every call should be logged with the person, tool and result. OpenTelemetry export is common; a per-person audit log is not.

Deployment and license: self-host effort, license terms, managed option

Compare how it runs (single binary, Docker Compose, Helm on Kubernetes), what the license allows, and whether there is a managed option when you no longer want to operate it.

Coverage: catalog size, remote and custom servers

Most open-source gateways ship no catalog, so you register every server yourself. A maintained catalog saves weeks; support for your own internal servers matters either way.

Run this checklist during a proof of concept:

1. Sign in with your identity provider, then remove a test user and confirm access is gone.
2. Connect the same app as two people and confirm each call uses that person's own token.
3. Restrict one group to read-only tools and try a write.
4. Find one tool call in the logs by person and time.
5. Add an internal server next to a public one.
6. Upgrade the gateway to a new version and check nothing breaks.

The top picks, compared

Values come from each project's own repo and docs; Not published means the project doesn't say. GitHub stars as of October 19, 2026.

The picks, one by one

1. Metorial

Metorial is an open-source MCP and skills gateway: one URL behind your company SSO that gives every employee and agent the tools their role allows, with 1,100+ MCP servers and 20,000+ tools in its catalog.

Best for: teams that want to self-host a gateway for the whole company, not only for developers.

Strengths

• SSO and SCIM, with every tool call running on the person's own OAuth connection through one Magic MCP URL.
• Access control by role down to single tools, a per-person audit log, and Protoguard prompt-injection checks on every call.
• A maintained catalog plus Skills that your experts write in the browser and sync to Git, so you aren't registering every server by hand.

Limits

• There is no LLM gateway or per-person model spend tracking yet, which Archestra, Bifrost and agentgateway include.
• Self-hosting runs the full platform with Docker Compose, which is more to operate than a single-binary gateway.

License · deployment · pricing: open core on GitHub; self-hosted with Docker Compose, Metorial Cloud or on-prem; free Dev plan, see pricing.

github.com/metorial/metorial · How the MCP gateway works

2. IBM ContextForge

ContextForge is an Apache-2.0 gateway and registry that federates MCP, A2A agents, REST and gRPC services behind one endpoint.

Best for: platform teams that want a permissively licensed gateway for MCP, agents and existing APIs.

Strengths

• Translates REST and gRPC services to MCP in the same gateway.
• SSO with Entra ID, Okta, Keycloak, Google and generic OIDC, with OAuth tokens scoped per user.
• 40+ plugins, including PII filtering, secrets detection and content moderation, and OpenTelemetry export.

Limits

• No hosted catalog: it ships a YAML list of public servers and you register your own.
• Request logs are per token, and security event logging is off by default.

License · deployment · pricing: Apache 2.0; PyPI, Docker or Helm on Kubernetes; free.

github.com/IBM/mcp-context-forge · Metorial vs ContextForge

3. Obot

Obot is an MIT-licensed MCP gateway, server host, skills registry and LLM gateway that you self-host on Kubernetes or use as Obot Cloud.

Best for: teams that want a broad open-source platform and have people to operate it.

Strengths

• An LLM gateway next to the MCP gateway, with per-user model access and cost analytics.
• Runs MCP servers in isolated environments with domain-level egress control.
• Device management inventories local AI clients, MCP servers and skills on employee machines.

Limits

• Entra, Okta and other enterprise identity providers need a free Community registration or an Enterprise license.
• Skills are authored in GitHub and read-only inside Obot.

License · deployment · pricing: MIT; Docker image to evaluate, Kubernetes for production, or Obot Cloud; free up to 100 users.

github.com/obot-platform/obot · Metorial vs Obot

4. Archestra

Archestra is an AGPL-3.0 platform that combines an MCP gateway and registry with an LLM proxy, agents, chat and guardrails.

Best for: teams that want hard LLM spend limits and MCP governance in one self-hosted product.

Strengths

• A curated catalog of 900+ evaluated MCP servers, behind one URL per gateway.
• LLM spend limits per organization, team, user or agent, blocking requests once a limit is hit.
• Deterministic session guardrails enforced outside the agent loop, plus SKILL.md skills edited in the UI.

Limits

• AGPL-3.0, with enterprise features under a separate license; SSO is an enterprise feature, free for companies under 30 users.
• Tool access is chosen per gateway rather than through per-tool allow and deny rules.

License · deployment · pricing: AGPL-3.0 with separately licensed enterprise files; Docker image or Helm; free, Enterprise through sales.

github.com/archestra-ai/archestra · Metorial vs Archestra

5. agentgateway

agentgateway is an Apache-2.0 proxy written in Rust for HTTP and gRPC, LLM, MCP and A2A traffic. It is a Linux Foundation project, and Solo.io sells an enterprise distribution.

Best for: platform teams that want one data plane for services, LLMs and MCP.

Strengths

• MCP authorization written in CEL over tools, prompts and resources, with identity-provider guides for Entra, Okta and Keycloak.
• Federates the tools of every target behind one MCP endpoint.
• An OpenAI-compatible LLM gateway and inline guardrails in the same proxy.

Limits

• Per-user OAuth doesn't work behind the single federated endpoint; servers needing it go on separate paths or use token exchange.
• Audit logs are structured access logs; user identity comes from JWT claims you add yourself.

License · deployment · pricing: Apache 2.0; binary, Docker or Helm on Kubernetes; free, Solo Enterprise through sales.

agentgateway.dev

6. Docker MCP Gateway

Docker's MIT-licensed gateway runs each MCP server in its own container and draws on a catalog of 300+ servers.

Best for: a developer or small team that wants each MCP server isolated in a container.

Strengths

• Signed images, SBOMs, secret blocking and network blocking for local servers.
• Ships with the MCP Toolkit in Docker Desktop, with one-command client connect.
• Turn individual tools on or off per profile.

Limits

• Each developer authorizes OAuth on their own machine; there is no shared, per-person login in the open-source version.
• Identity-provider sign-in and per-person audit logs are in the sales-led Enterprise Gateway.

License · deployment · pricing: MIT; Docker CLI plugin, included in Docker Desktop; free.

github.com/docker/mcp-gateway · Metorial vs Docker

7. Bifrost

Bifrost, from Maxim AI, is an Apache-2.0 AI gateway written in Go: OpenAI-compatible LLM routing with a built-in MCP gateway.

Best for: teams that already need an LLM gateway and want MCP in the same place.

Strengths

• Virtual keys carry budgets, rate limits and MCP tool allow-lists in one identity.
• Per-user OAuth to upstream MCP servers is in the free version.
• MCP Code Mode lets the model orchestrate many tools in a sandbox instead of loading every definition.

Limits

• SSO, SCIM, role-based access, audit logs and log export are Enterprise-only.
• MCP is one module of an LLM gateway, not the product's focus.

License · deployment · pricing: Apache 2.0; Docker, NPX, Go binary or Helm; free, Enterprise custom.

github.com/maximhq/bifrost

8. Kuadrant MCP Gateway

An Apache-2.0, Envoy-based MCP gateway for Kubernetes Gateway API and Istio, with Kuadrant auth and rate-limit policies.

Best for: platform teams already running Istio and the Gateway API.

Strengths

• MCP servers are registered as Kubernetes resources, reusing existing ingress, policy and rate limiting.
• Tool, prompt and resource access driven by identity-provider role claims.
• A per-user audit log of who called which tool, and a supported path through Red Hat Connectivity Link.

Limits

• Kubernetes only, and the repo's own overview still describes it as an early preview.
• Per-user upstream credentials rely on HashiCorp Vault rather than built-in OAuth.

License · deployment · pricing: Apache 2.0; Helm on Kubernetes or the OpenShift operator; free.

github.com/Kuadrant/mcp-gateway

9. Microsoft MCP Gateway

An MIT-licensed reverse proxy and management layer for MCP servers on Kubernetes, authenticating with Microsoft Entra ID.

Best for: teams standardized on Azure and Entra ID.

Strengths

• Entra ID bearer tokens with app roles checked per adapter or tool.
• Runs on local Kubernetes or AKS, next to Azure API Management for REST APIs.
• No new vendor for Microsoft-standardized companies.

Limits

• No integration catalog; you deploy and register your own servers.
• Per-user OAuth to upstream services is partial and depends on Azure's credential manager.

License · deployment · pricing: MIT; Kubernetes manifests on local Kubernetes or AKS; free.

github.com/microsoft/mcp-gateway · Metorial vs Microsoft MCP Gateway

10. Lunar MCPX

Lunar MCPX is a self-hosted MCP gateway with an MIT core and a per-seat enterprise tier. Boomi acquired Lunar.dev in July 2026, and MCPX is now part of Boomi's AI Gateway.

Best for: companies already on Boomi that want MCP and API traffic under one policy layer.

Strengths

• Aggregates local and remote MCP servers behind one endpoint, with per-agent tool controls.
• Pairs with an LLM and API gateway that shares the same policy and identity layer.
• Vetting and a sandbox for MCP servers before they reach production.

Limits

• SSO, per-user auth and full audit trails are in the Enterprise version.
• Its docs now redirect to Boomi pages, so the open-source project's roadmap is unclear.

License · deployment · pricing: MIT; Docker or Helm; free up to 50 users, Pro and Enterprise through sales.

github.com/TheLunarCompany/lunar

11. MCPJungle

MCPJungle is an MPL-2.0 gateway and registry in one Go binary, with tool groups and per-client access tokens.

Best for: a small team that wants a light, free registry without Kubernetes.

Strengths

• A single binary or Docker image, with SQLite locally and Postgres for teams.
• Runs stdio servers as child processes inside the gateway.
• Free under MPL-2.0, with no paid tier gating its access controls.

Limits

• No SSO or OAuth for clients connecting to the gateway yet; they use static tokens.
• Upstream OAuth is shared across the whole gateway, so there is no per-user login, and audit logging is limited.

License · deployment · pricing: MPL-2.0; Docker Compose or binary; free.

github.com/mcpjungle/MCPJungle

Which one should you choose?

If you are...	Your priority	Pick
An IT or platform team rolling AI out company-wide	SSO, per-person audit, a catalog and skills	Metorial
A platform team that wants a permissive license	Apache 2.0, MCP plus A2A and REST	IBM ContextForge
A team that wants one broad platform to operate	MCP, LLM and skills in one place	Obot or Archestra
A team already running Istio or many proxies	One data plane for every kind of traffic	agentgateway or Kuadrant
A solo developer running Claude Code locally	Zero setup, container isolation	Docker MCP Gateway
A small team that wants something light	One binary, no Kubernetes	MCPJungle

Whichever you pick, test it with a few real apps such as GitHub, Linear and Slack, and run the checklist above before rolling it out.

Related: best enterprise MCP gateways · what is an MCP gateway · Composio alternatives

Frequently asked questions

What is an open-source MCP gateway?

It is an MCP gateway whose code is public, so you can run it on your own infrastructure. It sits between AI clients such as Claude, Cursor and ChatGPT and the MCP servers they use, and handles routing, sign-in, access rules and logging in one place.

Which open-source MCP gateway is best for a company?

Metorial, because it is built for whole-company rollouts: people sign in with SSO, every tool call runs with their own OAuth connection and is logged against them, and teams can share skills. IBM ContextForge, Obot and Archestra also support SSO and per-user OAuth but leave more of the rollout to you.

Which open-source MCP gateway is easiest to run?

Docker MCP Gateway runs as a Docker CLI plugin and ships with Docker Desktop, and MCPJungle is a single binary or Docker image. Both suit one developer or a small team; neither offers SSO in its open-source version.

Do open-source MCP gateways support SSO?

IBM ContextForge, agentgateway and Kuadrant include SSO in their open-source code, and Metorial supports SSO and SCIM. In others, such as Docker MCP Gateway, Bifrost and Lunar MCPX, SSO is part of a paid enterprise edition, and MCPJungle does not support it yet.

Can I self-host an MCP gateway on Kubernetes?

Yes. IBM ContextForge, Obot, Archestra, agentgateway, Bifrost and Lunar MCPX publish Helm charts or Kubernetes deployments, Kuadrant and Microsoft MCP Gateway are Kubernetes-native, and Metorial self-hosts with Docker Compose.

Are open-source MCP gateways free?

The open-source editions are free to run on your own infrastructure. Several projects sell an enterprise edition or a managed cloud on top, and the real cost is the engineering time to operate, patch and secure the gateway.

Sources

1. Metorial competitor dataset (every competitor fact, with its source and quote)
2. Metorial on GitHub
3. IBM ContextForge on GitHub
4. Obot on GitHub
5. Archestra on GitHub
6. agentgateway docs
7. Docker MCP Gateway on GitHub
8. Bifrost on GitHub
9. Kuadrant MCP Gateway on GitHub
10. Microsoft MCP Gateway on GitHub
11. Lunar MCPX on GitHub
12. MCPJungle on GitHub