You connected one MCP server to your agent and it felt like magic. Then you connected a second, a third, and a fifth. Now every developer on your team has a different config file, three servers hold long-lived tokens in plain text, and nobody can say which tool call deleted that Jira ticket on Tuesday. That is the moment people start searching for an open source MCP gateway.
A gateway sits between your AI clients and your MCP servers. Agents talk to one endpoint. The gateway handles authentication, routing, rate limits, logging, and tool filtering. This article walks through the open source options, the managed alternatives, a realistic setup, and what each path costs.
💡 Quick answer: If you have a platform team and strict data rules, start with an open source gateway you host yourself. If you have no one to patch and monitor it, a managed gateway will usually cost less than the engineer hours you would spend.
What an MCP Gateway Does
The Model Context Protocol lets an LLM call external tools in a standard way: read a file, query a database, post to a chat channel. Each tool lives on an MCP server. Without a gateway, each client connects directly to each server.
A gateway replaces that web of connections with a single front door.

The Core Jobs
Most gateways, open source or not, handle the same short list:
- Routing: one URL for the agent, many upstream MCP servers behind it.
- Authentication: OAuth or token checks at the edge, so servers do not each reinvent login.
- Authorization: which user or team can call which tool.
- Rate limiting: stop a looping agent from hammering an internal API.
- Audit logs: who called what, with which arguments, and what came back.
- Tool filtering: hide risky tools such as "delete" or "write" from certain agents.
Why Not Connect Directly
Direct connections work fine for a solo developer. They break down when more than a handful of people share tools. Credentials get copied across laptops. Updating one server means touching every client config. And when something goes wrong, you have no single place to look.
Open Source MCP Gateway Options
The open source field moves fast, so treat any roundup as a snapshot. These are the projects that come up most often in 2026 comparisons.

IBM ContextForge
ContextForge is described as a free MCP gateway, registry, and proxy with an admin interface, OAuth-scoped authentication, and OpenTelemetry tracing. There is no license fee and no per-user or per-token billing, and there is no hosted version from the project itself. Your real cost is the infrastructure and the time to run it.
Docker MCP Gateway
Docker's approach runs each MCP server in its own container and exposes them through one gateway process. The appeal is isolation: a misbehaving server cannot reach your host. It fits teams that already live in Docker Desktop or Compose.
MetaMCP
MetaMCP aggregates several MCP servers into one endpoint and lets you group them into namespaces. It is popular with small teams who want a web UI to toggle tools on and off without editing JSON.
Obot
Obot is MIT licensed, with extra enterprise features unlocked by a license key. Its Docker quick start is meant for evaluation only, so plan on a proper deployment before real users touch it.
Microsoft's Kubernetes Gateway
Microsoft publishes an open source MCP gateway aimed at Kubernetes. It suits organizations that already run clusters and want session-aware routing for many servers.
Tyk and Open Core Models
Some projects use an open core model: the base is open, but the features you need for production, such as SSO or advanced policy, sit behind a paid tier. Check the license table before you build on it.
| Option | License model | Best fit | Hosted version |
|---|
| IBM ContextForge | Free, no license fee | Platform teams wanting a registry plus proxy | No |
| Docker MCP Gateway | Open source | Container-first teams | No |
| MetaMCP | Open source | Small teams wanting a UI | Community options |
| Obot | MIT plus enterprise license key | Teams that may need SSO later | Yes, paid |
| Microsoft MCP gateway | Open source | Kubernetes shops | No |
| Tyk | Open core | Teams already using Tyk for APIs | Yes, paid |
⚠️ Heads up: Roundups disagree on licensing for several projects, and some tables published online are garbled. Open the repository, read the LICENSE file, and confirm before you commit.
Managed Gateway Alternatives
If you would rather not patch a gateway at 2 a.m., a managed service takes over hosting, scaling, and upgrades. You pay in subscription or usage fees instead of engineer time.

Kong
Teams that already run Kong can use its MCP gateway as a feature of a tool they already own. It is part of the larger AI gateway offering and leans on paid plugins, so it reads as an enterprise product rather than a weekend experiment.
Cloudflare MCP Server Portals
Cloudflare offers a free plan that is enough for a proof of concept. Larger plans sit under Cloudflare One and are priced on request. If your staff already authenticate through Cloudflare Access, this is the shortest path to single sign-on for MCP.
Portkey
Portkey began as an LLM gateway and added MCP controls. Pricing is usage based with enterprise plans for stricter governance. It makes sense when you want model routing and tool routing in one place.
Composio
Composio ships a large catalog of hosted integrations. One published breakdown lists a free plan with 20,000 tool calls a month, a mid tier around $29 a month with 200,000 calls, and a business tier near $228 with 2,000,000 calls. Other roundups describe the tiers differently, so trust the live pricing page over any blog.
AWS Bedrock AgentCore Gateway
AWS bills by consumption: roughly $0.005 per 1,000 invocations, plus charges for search calls and indexed tools, with data transfer on top. There is no flat subscription, which suits spiky workloads and punishes chatty agents.
TrueFoundry and Similar Platforms
TrueFoundry lists a free developer plan with 50,000 requests a month, then paid tiers starting around $499 a month. Remember that most pricing commentary about competitors comes from other vendors, so read it with some salt.
Open Source vs Managed: Real Costs
Sticker price is the smallest part of the bill. Here is how the two routes compare when you count everything.

Open Source Cost Breakdown
| Cost item | Typical impact |
|---|
| License | $0 for most projects |
| Compute and storage | A small VM or a few pods, usually modest |
| Setup time | Days for a basic deployment, weeks for SSO and policy |
| Patching and upgrades | Ongoing, and someone must own it |
| Incident response | Your team, with community help at best |
| Compliance evidence | You build the audit trail yourself |
Managed Cost Breakdown
| Cost item | Typical impact |
|---|
| Subscription | Free tier to a few hundred dollars a month for small teams |
| Usage fees | Scales with tool calls or requests |
| Setup time | Hours to a couple of days |
| Patching and upgrades | Handled by the vendor |
| Incident response | Vendor SLA on paid plans |
| Compliance evidence | Often included on higher tiers |
A Simple Rule of Thumb
If your gateway handles under a few hundred thousand tool calls a month and you have no dedicated platform engineer, managed wins on total cost. If you push millions of calls, have regulated data, or already run Kubernetes with a team to match, open source usually wins within a year.
💡 Tip: Price the engineer hours honestly. Two days a month of maintenance at a loaded engineering rate often exceeds a mid-tier managed plan.
Setting Up a Gateway Step by Step
The exact commands differ per project, but the sequence barely changes. Here is a path that works for most self-hosted setups.

Step 1: Inventory Your Servers
List every MCP server your team uses today. For each one, write down who owns it, what credentials it holds, and which tools it exposes. Most teams find two or three servers nobody remembers installing.
Step 2: Pick a Deployment Target
A single VM with Docker Compose is fine for a pilot. Use Kubernetes only if you already run it. Put the gateway behind HTTPS from day one, since MCP traffic often carries sensitive data.
Step 3: Register Upstream Servers
Add each MCP server to the gateway's registry. Start read-only. Disable any tool that writes, deletes, or sends messages until you have tested it.
Step 4: Wire Up Authentication
Connect your identity provider through OAuth or OIDC. Avoid shared static tokens. If a project only supports static tokens at first, rotate them on a schedule and store them in a secrets manager, never in the repo.
Step 5: Point One Client at It
Change one agent or IDE to use the gateway URL and nothing else. Run your normal workflows for a week. Compare logs against what you expected.
Step 6: Add Limits and Alerts
Set per-user and per-tool rate limits. Alert on error spikes and on any call to a high-risk tool. Then roll the gateway out to the rest of the team.
A Minimal Compose Sketch
Most projects publish a container image. A pilot usually needs only a gateway container, a database for the registry, and a reverse proxy for TLS. Keep configuration in environment variables and secrets in your platform's secret store.
Security Checks Before Launch
A gateway concentrates risk. If it is compromised, an attacker reaches every tool behind it. Treat it like an identity provider, not a utility script.

Authentication Basics
Require strong identity for every connection. Prefer short-lived tokens issued through your identity provider.

Prompt Injection and Tool Poisoning
A malicious document can tell an agent to call a tool it should not. A gateway helps by enforcing an allowlist outside the model. If a tool is not on the list, the call fails no matter what the model was tricked into asking.
You can also place a safety classifier in front of risky calls. A model such as Llama Guard 4 12B is built for content moderation and can screen inputs and outputs before they reach a tool.
Checklist Before Going Live
- Every upstream server sits on a private network, reachable only through the gateway.
- Write and delete tools are off by default.
- Logs go to a store the gateway cannot modify.
- Secrets live in a vault, not in config files.
- You have a documented way to revoke a user in minutes.
Monitoring and Day Two Work
Launching is the easy part. A gateway earns its keep when something breaks and you can see why in two minutes.

What to Track
Watch request count, error rate, latency per tool, and token or cost estimates per team. OpenTelemetry support, which ContextForge advertises, lets you send traces to the same system you use for the rest of your services.
Choosing the Model Behind the Agent
The gateway does not care which LLM calls it, but results do. Tool calling quality varies between models, and long multi-step runs punish weaker ones. For agent work, these are worth testing against your own tools:
Run the same ten tool-calling tasks through each and compare success rate, number of retries, and cost. The gateway logs make that comparison easy.
Who Owns It
Name one owner. Gateways fail quietly when everyone assumes someone else is watching them.
How to Choose Your Path
Match the option to your situation, not to the loudest blog post.

| Your situation | Best starting point |
|---|
| Solo developer or tiny team | MetaMCP or a managed free tier |
| Container-first team, no Kubernetes | Docker MCP Gateway |
| Platform team wanting registry and tracing | IBM ContextForge |
| Kubernetes shop | Microsoft's gateway |
| Already on Kong or Cloudflare | Their built-in MCP features |
| Regulated data, need vendor SLA | Managed enterprise plan |
| Spiky usage, no ops staff | Consumption-billed cloud gateway |
Three Common Mistakes
- Starting with every server. Begin with two or three, prove it works, then expand.
- Skipping the owner. An unowned gateway becomes a liability within months.
- Trusting a pricing blog. Numbers change monthly and sources conflict. Always read the vendor page.
Try It on Picasso IA
An MCP gateway is only useful if the agents behind it are good at using tools. That is where model choice matters, and where hands-on testing beats reading benchmarks. Picasso IA gives you one place to try leading language models side by side, from Claude Sonnet 5 and GPT 5.6 Sol to open weight options like Granite 4.1 8B.
Paste in the same tool-use prompt, compare how each model plans and answers, and pick the one that fits your gateway workload. When you are ready to go beyond text, browse everything the platform offers, including image, video, and audio models, at picassoia.com/en/all-models. Open a model, run a few prompts, and see what your next project can do.