MCP security audit.
Our MCP security audit checks your MCP servers, and the agents that use them, against the MCP specification and the attacks it warns about. You get every finding ranked, with the code to fix it.
- Paid audit
- Starts with a free 30-minute intro call
- Fixes as code
What the audit covers.
An MCP server lets an AI model call your tools. The audit checks who can call them, with which token, and what the tools do with what the model sends.
- Authorization
- The OAuth flow the MCP specification sets out: protected resource metadata, the 401 challenge, and tokens checked for signature, issuer, expiry and audience. No token passed on to another API.
- Scopes and tool access
- Which scope each tool needs, checked when the tool runs rather than only at connect, and whether tools that change or delete anything ask a person first.
- What tools do with input
- Tool handlers read for the classic attacks: any URL fetched on request, file paths outside their folder, shell commands and raw SQL built from the model’s input.
- What tools tell the model
- Tool descriptions, schemas and results checked for hidden instructions aimed at the model, and annotations checked against what each tool really does.
- Transport and deployment
- Origin checks, local servers bound to localhost, session handling on servers that still use sessions, rate limits, and secrets kept out of logs, errors and URLs.
- Agents and clients
- Which MCP servers your agents and your team connect to, which tokens they hold, and whether one tool’s output can steer a call to another tool with more access.
The risks behind each area are explained in our MCP security best practices, and the sign-in flow in how to add OAuth to an MCP server.
What you get.
A report your engineers can work from, not a list of warnings to interpret.
-
Every finding, ranked
What is wrong, where in the code or configuration, how it could be used and how serious it is, so what exposes data or allows changes gets fixed first.
-
Fixes as code
Code for your server, ready-made for the official TypeScript and Python SDKs and FastMCP, written to your own code for anything else.
-
A tool-by-tool access map
Each tool with the scope it should need and whether a person should approve it, ready to turn into roles.
How the audit works.
Scope is agreed on the intro call, so you know what will be checked before you commit.
-
Step 1 · Free Intro call
A 30-minute call about your MCP servers and agents, and which of them the audit should cover.
-
Step 2 Outside checks
The server is checked from outside the way an MCP client or an attacker would see it.
-
Step 3 Code and configuration review
A review of the server’s code, its authorization server settings and the agents that call it, with the access you choose to grant.
-
Step 4 Report
The ranked findings, the code fixes and the tool-by-tool access map.
Building the agents too? See our custom AI agent development. If your product’s own sign-in and roles need the same care, see B2B SaaS authentication and SSO.
Free check or full audit.
Start with the free checker. Book the audit when you need what it can’t see.
-
MCP Security Checker
- Public servers, checked from outside
- Metadata, the 401 challenge and how bad tokens are refused
- Tool review when the server is open, or from a pasted tools/list
- Code fixes for the official SDKs and FastMCP
-
MCP security audit
- Public and private servers, from outside and from the code
- How real tokens, audiences and scopes are checked
- What each tool does with its input
- The agents and clients that connect, and the tokens they hold
Questions, answered.
Anything else, ask on the intro call.
What is an MCP security audit?
A review of an MCP server, and of the agents and clients that use it, against the MCP specification and the attacks it warns about: missing or weak authorization, tokens passed through, tools with too much access, unsafe tool input, and instructions hidden in what tools return.
How is it different from the free MCP Security Checker?
The checker sees only what a public server shows from outside: its metadata, how it refuses bad tokens, and its tools if it is open to everyone. The audit also reads your code and configuration, so it covers how real tokens and scopes are checked, what each tool does with its input, and the agents that connect.
Can you audit MCP servers we use but didn’t build?
Yes, from your side: which servers your team and agents connect to, what those servers can reach, and which tokens and permissions you have given them. Their own code is reviewed only where you can share it, such as open-source servers.
Which languages and SDKs do you cover?
Any. Code fixes come ready-made for the official TypeScript and Python SDKs and FastMCP, and are written to your code for other languages and frameworks.
How much does it cost, and how long does it take?
Both depend on the scope: how many servers, tools and agents the audit covers. Scope and cost are agreed after the free intro call, before any paid work starts.
What access do you need?
Read access to the code in scope, and a test or staging deployment with a test account where possible. You decide what access to grant.
Can you fix what you find?
Yes. Your team can fix the findings from the report, or NestedBit can make the fixes as a fixed-scope pilot.
Let agents in, safely.
Start with a free 30-minute intro call, or email hello@nestedbit.com.