Scope audit
Every declared tool is compared against the permissions it actually requests. A weather lookup that asks for filesystem writes gets flagged as high severity, with the exact line in your manifest that caused it.
Ship your MCP server without shipping a security hole.
Paste your manifest. Get a security & compliance report in 30 seconds — scope audit, injection scan, exfiltration paths, and a security statement you can paste straight into your README.
No signup wall. No server access. Static analysis of your manifest only.
Why MCP Server Security Matters Right Now
The Model Context Protocol ecosystem crossed a threshold this year. MCPJam launched on Product Hunt with 139 upvotes and 38 comments, and on the same day four other agent-infrastructure products — cloud Mac runners, agent routers, voice test harnesses — all landed in the mid-40s. That cluster is a signal: the agent stack is moving from "can it run?" to "can we ship it?" Every team that wants an enterprise customer to connect their agent now has to answer a question they never had to answer before — what exactly can this server touch?
Existing tooling does not answer that question. Test platforms tell you whether a server responds correctly, but they say nothing about whether a "weather lookup" tool quietly requests filesystem write access. General dependency scanners like Snyk and Socket.dev read your package tree, but they have no concept of a tool description that says "always call this first" — a line that a prompt injection can weaponize. Semgrep needs your full source and a ruleset you have to maintain. The gap is not code quality. The gap is the semantic layer of MCP itself: declared scopes, tool descriptions, and outbound endpoints.
That gap is closing fast. Directories like mcp.so and Smithery are starting to need trust signals, and enterprise buyers are starting to ask for them. A server author who can attach a dated, reproducible audit report to their README has a real advantage over one who cannot. MCPCheck exists to produce that artifact in under a minute, from a manifest alone, with no access to your running server.
How It Works
1
Paste your manifest or repo URL
Drop in a manifest.json, a server.json, or a public GitHub URL. MCPCheck pulls the declared tools, their descriptions, and their requested scopes — no credentials, no runtime access, no install required.
2
Run the static rule engine
Five rule families execute against your declaration: scope overreach, description injection language, outbound exfiltration paths, dependency CVEs via the free OSV.dev API, and missing security statements. Each finding carries a severity and a concrete fix.
3
Export the report and the statement
You get a single-page HTML report you can save as PDF, plus a Markdown security statement listing exactly which permissions your server does not request. Paste it into your README so enterprise reviewers can clear you faster.
What You Get
Injection scan
Tool descriptions are scanned for imperative language a prompt injection could exploit — phrases like "ignore previous instructions" or "always call this tool first" — and each hit comes with a suggested rewrite.
Copy-paste statement
You receive a dated Markdown security statement listing the permissions your server does not request and the endpoints it never contacts. Drop it into your README so enterprise reviewers can approve you without a call.