MCP servers connect AI agents to real tools and data. An agent can look up an order, search your files or update a record, all through the server your team built. That makes launching a new server, or pushing an update to one that's already live, a high-stakes moment.
Most teams go into that moment with a feeling rather than an answer. It worked on someone's laptop, a few manual tries looked fine, and the release goes out. This post covers what worries teams most, and how Guardian's MCP scan turns "it seems fine" into something you can see and share.
What keeps teams up at night
When teams talk to us about shipping an MCP server, the same fears come up again and again.
It connects in dev but fails in production
Development and production are rarely identical. Different addresses, different sign-in, different network rules. A server that connected perfectly while being built can fail the first time a real agent tries to reach it.
A tool that worked yesterday breaks after an update
Every change is a chance to break something that used to work. A renamed field or a tweaked description can quietly turn a reliable tool into a broken one, and nobody notices until an agent hits it.
The AI picks the wrong tool or sends the wrong values
AI agents don't read your documentation the way a developer does. They rely on what each tool says about itself. If those descriptions are vague, an agent may choose the wrong tool, or send a value the tool never expected.
Tools that change data can do damage
Some tools create, update or delete things. Used wrongly, they do real harm, and that harm is much harder to undo than a failed request.
It leaks credentials or accepts connections from anyone
A server that accepts connections without sign-in, runs over an unprotected connection, or has secrets sitting in its launch settings is a risk to every system behind it.
Nobody can show a customer or auditor that it was checked
Even when a team did check carefully, the proof is usually a chat message or someone's memory. When a customer or auditor asks how you know the server is ready, there's nothing to point to.
Four categories, one score
Guardian's MCP scan answers those worries with four questions. Each one is its own category, and each is made up of specific checks.
Integration: "Does the system respond?"
This is the basics. The scan checks:
- Connection: can it reach the server at all?
- Discovery: does the server tell an agent what tools it offers?
- Protocol: does the server speak MCP correctly?
Unit: "Does each tool work?"
Next, the scan looks at the tools one by one:
- Tool definitions: is each tool described clearly and completely?
- Tool execution: when the scan uses a tool, does it work?
Evaluation: "Can an AI use it well?"
A server can work perfectly and still be hard for an AI to use. This category checks:
- Tool selection: given a task, does the AI pick the right tool?
- Parameters: does it send the right values?
- Answer quality: is the result useful?
- Multi-step use: can it combine several tools to finish a larger job?
Security: "Is it safe to connect to?"
Last, the scan looks at whether connecting is safe:
- Transport: is the connection protected?
- Authentication: does the server ask who is connecting?
- Tool safety: are the tools safe to hand to an agent?
- Launch config: if the server is started locally, are its launch settings safe? This check only appears for servers started that way.
What you get back
The four categories roll up into a single score out of 100. You also see a bar for every check, so you can tell exactly where points were lost instead of just seeing a number. For each problem, the scan lists a finding with how to fix it, so the next step is clear.
When to run it
Run it before the first launch, so you know where you stand. Then run it after every change, because a server that was fine last week may not be fine today.
It works alongside testing on your own machine. Our post on testing your MCP server before it leaves your laptop covers catching problems while you build. The scan is the check on the server as an agent would see it.
Try it on your server
You can run a scan any time at /tools/scan. The result is a record you can share with your team, and a clear list of what to fix.
A scan gives you evidence, not a guarantee. It shows what was checked and what was found at that moment, which is a far better starting point than a hunch.
One reminder: only scan servers you own or have permission to test.
Photo by Tony Marinescu on Unsplash.
