
The first question most security teams ask about an enterprise MCP gateway isn’t what it does. It’s where it runs. If every agent’s tool calls, credentials, and logs pass through one control point, that control point had better sit somewhere your data is allowed to go.
For some organizations, a managed service is fine. For others, especially in regulated industries, the gateway has to run inside their own cloud account, data center, or network. This guide explains what a self-hosted MCP gateway is, the deployment models to choose from, what each one costs you in operational effort, and how to decide.
What is a self-hosted MCP gateway?
A self-hosted MCP gateway is an MCP gateway that runs in infrastructure your organization controls, such as your own cloud account, Kubernetes cluster, or data center, so AI tool traffic, credentials, and logs stay inside your environment instead of passing through a vendor’s shared service.
An MCP gateway sits between AI agents and the MCP servers they use. It checks who’s calling, decides which tools they can use, applies policy to each call, and records what happened. Self-hosting changes where that happens, not what it does.
Why enterprises self-host an MCP gateway
The reasons are usually concrete, and they’re worth naming before you pick a model.
- Data residency. Tool inputs and outputs can contain customer records, source code, or financial data. Some of that can’t leave a country, a cloud account, or a network.
- Regulated workloads. Financial services, healthcare, and public sector teams often need AI traffic to stay inside environments that have already passed their own audits.
- Reaching internal systems. Many MCP servers front internal databases and applications that aren’t exposed to the internet. A gateway inside the network can reach them without opening new paths in.
- Credential custody. The gateway often holds or brokers the credentials agents use. Some teams require those secrets to stay in their own vault.
- Log ownership. Every tool call log is sensitive. Keeping logs in your own environment makes it easier to feed your SIEM and meet retention rules.
The tradeoff: control vs. operational work
Self-hosting gives you control. It also hands you a job. Before choosing it, be honest about who will:
- Run it. Provision, scale, and monitor the gateway, and keep it highly available, because when it’s down, agents can’t use their tools.
- Secure it. Harden the gateway itself. It’s now one of the most sensitive systems you operate.
- Patch and upgrade it. Apply security fixes quickly and keep up with new releases.
- Keep up with MCP. The protocol is still changing. The MCP specification has gone through several revisions, including a new HTTP transport and changes to authorization, and your gateway has to keep up.
- Maintain connectors. Someone has to keep MCP server connections, credentials, and configurations current as the systems behind them change.
None of that is a reason not to self-host. It’s a reason to plan staffing and ownership before you do.
MCP gateway deployment models compared
| Model | Where it runs | What stays in your environment | Operational effort | Best fit |
|---|---|---|---|---|
| Managed SaaS | The vendor’s cloud | Nothing by default. Traffic passes through the vendor’s service | Lowest | Fast rollout, no strict residency rules |
| Private cloud | Your own cloud account | Traffic, credentials, and logs, inside your cloud boundary | Medium | Teams with residency or audit needs who run cloud infrastructure today |
| On-premises | Your data center or private Kubernetes | Everything, inside your network | Highest | Sovereign, air-gapped, or highly regulated environments |
| Hybrid | Control functions in one place, execution in another | Execution and data in your environment, management elsewhere | Medium to high | Gradual migrations, or central management with local execution |
Hybrid deserves a closer look
Hybrid deployments split the gateway into two parts. The control plane handles configuration, policy, and administration. The execution side handles live traffic: the actual tool calls and the data in them. Running execution inside your environment while managing it centrally can give you most of the data control of self-hosting with less of the operational load. The key questions are what data, if any, crosses back to the control plane, and what happens to running agents if the connection between the two drops.
What should stay the same in every model
Whichever model you choose, the controls themselves shouldn’t change. A self-hosted gateway that enforces weaker policy than a managed one isn’t a security upgrade.
- The same identity. Connect to your identity provider so the gateway knows which person every agent is acting for.
- The same policy. Tool-level and parameter-level access, approval for high-impact actions, and rate limits should work identically everywhere.
- The same inspection. Tool definitions should be scanned before they’re exposed, and inputs and outputs checked as they flow.
- The same logs. Every call, including denied ones, should land in one place your security team can search.
If you run more than one model, for example SaaS for most teams and on-premises for one regulated business unit, consistent policy across both matters even more.
How to decide
Work through these questions in order.
- Where must the data stay? Map which tool inputs and outputs fall under residency, regulatory, or contractual limits. This usually decides the model.
- What does the gateway need to reach? If key MCP servers sit inside your network, a gateway inside the network may be simpler and safer.
- Who will operate it? Name the team that will run, secure, patch, and upgrade the gateway. If there isn’t one, a managed or hybrid model may be the better fit.
- What are your availability needs? Decide how much downtime agents can tolerate and design for it.
- How will logs reach your security tools? Plan the path from gateway logs to your SIEM before go-live.
- Is policy portable? Make sure the same rules apply if you move between models later.
Rolling out a self-hosted MCP gateway
- Start with an inventory. Find the MCP servers already in use, including local ones on developer machines.
- Pilot with one team. Deploy in a non-production environment and connect a small set of approved servers.
- Integrate identity and logging first. Connect your identity provider and SIEM before adding more users.
- Set policy before scale. Define role-based tool access and approval rules while the user count is small.
- Load test and plan for failure. Confirm the gateway handles peak agent traffic and fails safely.
- Expand by risk. Move the highest-risk MCP connections onto the gateway first, then the rest.
How Airia approaches deployment
Airia offers more than one way to run its platform. According to Airia’s deployment options, the platform is available as fully managed SaaS, in private cloud environments on major cloud providers, on-premises, and in hybrid configurations where control plane components run in one environment and execution components run in another.
The Airia MCP Gateway provides the core controls described above: admin approval before servers can be used, access at the app, tool, and parameter level by role, scanning of tool names and descriptions for hidden prompt injection before they’re exposed, and a searchable log of every tool call, whether it succeeded, was denied, or errored. Airia Discover helps with the first step of any rollout by finding the MCP servers already running across your environment.
Deployment details vary by environment and requirement, so the best next step is a conversation about where your data needs to stay.
Conclusion
Self-hosting an MCP gateway is a decision about data and ownership, not features. Start with where your data must stay and what the gateway needs to reach. Be clear about who will run it. Then make sure identity, policy, inspection, and logging work the same way in every environment you choose. Get those right, and the gateway can live wherever your requirements say it should.
Decide where your MCP gateway should run, with the facts in front of you. Request a demo to talk through your data residency needs and see which Airia deployment model fits.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.