Airia's MCP Gateway Now Routes the Right Tools to the Right People, Automatically

The problem with MCP at scale is not the protocol. It is everything that comes after.

MCP has become the uncontested standard for connecting AI agents to enterprise tools. Organizations are not debating whether to adopt it. They are living with what adoption actually looks like: dozens of MCP servers, hundreds of tools, and a configuration problem that multiplies every time a new user, client, or application enters the picture.

The token problem hits first. Every tool connected to an MCP server sends its full description into the model’s context window on every call. An organization with 200 available tools pays for all 200 on every interaction, whether the user needed two of them or twenty. That cost accumulates before a single productive task is completed. In practice, organizations running large tool catalogs have found themselves loading 250 or more tool definitions into every single call, with context window overload and runaway token spend as the immediate result.

The accuracy problem follows. Every tool description loaded into the context window is noise competing with the actual query. The more tools in the window, the harder it is for the model to act on what the user actually asked. Models perform measurably worse with crowded context windows, not because the tools are wrong, but because there is simply too much competing for attention.

Then there is the deployment problem. API specifications that convert into MCP servers often produce hundreds or thousands of individual tool definitions. A single API catalog can generate an MCP server with 1,500 endpoints. Without a filtering layer, connecting a catalog that size to a model is not a realistic option. The context window fills before the agent can do anything useful, and the entire configuration becomes too unwieldy to ship. The standard workaround has been manual curation: pick the most important endpoints, leave the rest on the floor. That means capability gets sacrificed before deployment even starts.

And underneath all of it is the operations problem. Every developer today has to configure their own MCP servers individually. Every new user who needs access to a subset of tools requires a configuration built and maintained by someone in IT. Rolling a new tool out to 200 engineers means touching 200 devices. That overhead does not shrink as MCP deployments grow. It compounds.

This release is built around all four of those problems.

The gateway is not just a security control. It is how AI capabilities reach people.

Before getting into what shipped, it is worth naming something that gets lost in how MCP is typically discussed. A gateway is not only a security tool. It is the distribution layer that determines how AI capabilities move through an organization. It is how a new tool gets to 500 engineers with a single configuration change. It is how an AI operations team maintains consistency across every client, every user, and every role without rebuilding configuration every time something changes. Security and enablement are the same motion. This release serves both.

What shipped

Airia Radar

Radar is the intelligence layer that makes large-scale MCP deployment viable. The core mechanism is straightforward. Instead of loading every available tool description into the context window on each call, Radar introduces three lightweight native tools as the only defaults: Search, Execute, and Manage Cache.

When an agent issues a query, it calls Radar’s Search function. Search runs a semantic match across every tool available in the gateway, regardless of how many tools the gateway holds, and returns an ordered list of the most relevant options for that specific query. The agent selects the tools it needs and executes them. Those tools get added to the cache automatically. On every subsequent call that requires the same tools, the agent retrieves them from cache instantly without searching again.

The cache builds over time based on actual usage patterns. Tools each user reaches for most surface faster the longer they work through the gateway. There is no pre-configuration required, no manual curation of which tools to include, and no assumption made upfront about what any user will need. The gateway learns.

This matters for more than token cost. Every tool description that does not load into the context window is noise that does not compete with the actual query. Models work with cleaner signal. Responses improve not because the underlying model changed but because the context it receives is sharper.

The deployment implication is even more significant. An API catalog that generates an MCP server with 1,500 endpoints is fully usable through Radar. Previously, deploying a server that size meant choosing between manual curation and context window overload. With Radar, organizations connect the full catalog, and Radar manages relevance automatically. Every endpoint stays accessible. None of them permanently occupy space in the context window. Tool sets that were previously too large to deploy at all become viable on day one.

Airia absorbs the cost of every Radar call internally. Organizations see none of it on their token bill. What they see instead is context windows that carry only what each call actually requires, and token spend that reflects productive work rather than overhead.

Airia’s MCP catalog now includes nearly 2,000 pre-configured servers, giving organizations immediate access to a broad range of tools and data sources without custom setup.

Dynamic MCP Gateway

One gateway address serves every user in the organization with a personalized tool set. Admins assign tool groups to roles once. Users control their own preferences within those boundaries through a self-service dashboard called Your Apps, with no IT involvement required for individual changes. Every major MCP client including Claude Code, Cursor, and Codex is supported, with plain-language setup instructions and deep links available directly from the dashboard.

Organizations using the Airia Endpoint Agent can push the Dynamic Gateway to users automatically. Airia’s AI Discovery capability takes this further. When Discovery identifies existing MCP client configurations running on devices across the organization, it replaces those local server setups with the Dynamic Gateway automatically. A new employee opens their AI client on day one and the gateway is already there, pre-configured with the tools their role requires. No setup. No instructions. No configuration for anyone to get wrong.

Skills over MCP

Airia now supports the formal MCP protocol extension for Skills. Organizations connect public Skills repositories or private authenticated ones directly to the gateway. Skills appear alongside tools in any MCP-compliant client and execute natively with no additional configuration required.

Tool Annotations

Every tool in the Airia MCP catalog carries a semantic annotation derived from what the tool actually does: read only, destructive, or otherwise. Admins filter by annotation when building gateway configurations. Creating a read-only environment or excluding destructive tools no longer requires reviewing every tool description individually. Every tool also carries a scan status confirming it has been evaluated for prompt injection risks before it reaches users.

Granular access controls

Admins assign access at the app level, the tool level, and the parameter level by role, giving progressive control over exactly what each user and group can access and do within the gateway.

Why the combination matters more than any individual capability

Radar alone reduces token cost and improves model accuracy. Custom MCP servers give organizations control over exactly which tools and endpoints they expose. The ability to convert any OpenAPI specification directly into a functioning MCP server means organizations do not need custom development to bring their existing APIs into the ecosystem. Each of these capabilities delivers value on its own.

Combined, they unlock something none of them can do individually. An organization takes an existing API specification, converts it into an MCP server with however many endpoints the API contains, connects it to the gateway, and lets Radar handle relevance. No curation. No compromise on capability. No context window problems. The full API surface is available to every agent that needs it, and only the tools each specific query requires ever load into the context window.

That is not a feature. That is a different category of what enterprise MCP deployment can look like.

Who this is for

Security and IT leaders managing MCP deployments across a growing user base who need a control plane that scales with them, not against them. CIOs and AI operations teams who need to distribute tools to engineers and end users without rebuilding configuration every time something changes. Platform teams spending too much time on gateway setup and maintenance. And the developers and knowledge workers using AI clients every day who want the right tools available without managing MCP servers themselves.

Organizations already running a gateway through another provider are not blocked from any of this. The AI Security Gateway and MCP Gateway sit alongside existing tooling in the pipeline, delivering observability, cost controls, and MCP-layer capability that other providers do not offer.

The Dynamic MCP Gateway enhancements are available now for all Airia customers. Learn more.