Check what your AI agents and MCP servers expose, before you ship

The problem

Model Context Protocol (MCP) servers are becoming a common way to give AI agents access to tools and data, but it’s easy to stand one up without authentication, especially during prototyping, and then forget to add it before the endpoint becomes reachable from the public internet. A public MCP server with no auth can be called by anyone who finds it, not just the agent it was built for.

The problem runs in both directions. Agents that call out to unfamiliar tools or endpoints generally have no built-in way to check whether that endpoint is what it claims to be before sending a request. And when a tool sends a response back, that response is just more text the model reads, which means it can carry prompt-injection content: instructions hidden in the data that try to steer the model’s behavior, with nothing in most setups filtering for that today.

On top of that, many agent setups keep no audit log of what the agent actually called, when, or with what result, which makes it hard to reconstruct what happened after the fact if something goes wrong.

What you can do for free today

• Check if your MCP endpoint is reachable and whether it demands auth — a quick way to confirm your server isn’t open to the public without authentication.

For background on securing an MCP server more generally, see the how to secure an MCP server guide and the MCP developer reference.

What AdminKit will add

AdminKit’s MCP Gateway is in early access and development. The plan is to sit in front of an MCP server and add OAuth-based authentication, tool allowlists, filtering of tool responses for injection content, and an audit log of what was actually called. Alongside it, the planned Agent Trust API is meant to let an agent ask a quick trust question about an unfamiliar endpoint before calling it, rather than calling first and finding out later. Neither product is live yet; the free mcp-check tool is what’s available today, and it reports reachability and auth status for a single endpoint rather than ongoing protection.

Together, the two planned layers are meant to cover both directions of the problem: the Gateway protects a server you run from being called by the wrong caller or returning unfiltered responses, and the Trust API helps an agent you build decide whether an endpoint is worth calling in the first place. Until those ship, the free mcp-check tool and the guide above are the most useful starting points for checking your own MCP setup.