Loading page
About this tool
Compare MCP tool manifests locally and highlight added capabilities, changed schemas, and risky permission patterns before upgrading.
The MCP Permission Diff Auditor compares two Model Context Protocol tool manifests side by side and reports which tools were added, removed or changed, then flags each addition and change against seven keyword risk rules covering destructive actions, code execution, external writes, filesystem writes, network access, sensitive-data access and broad reads. It is for anyone about to upgrade an MCP server and wanting to know what new capabilities the new version quietly grants. Both manifests are parsed and diffed in your browser — nothing is uploaded and no MCP server is contacted or executed.
Open MCP Permission Diff Auditor on AltFTool — it loads instantly in your browser.
Paste the installed JSON into Previous manifest and the upgrade into Proposed manifest, or press Load safe demo for the sample pair.
Press Swap versions if they are the wrong way round — the diff normalises names, descriptions, input schemas and annotations before comparing.
Read the Added, Removed, Changed, High risk and Unchanged counts with each tool's risk badge, then Copy report or Download report as mcp-permission-diff-audit.txt.
It normalises names, descriptions, input schemas and annotations, so reordered or reformatted JSON does not show up as a change.
Changed tools are described as description changes, inputs added or removed, new required inputs, or annotation/permission changes.
It walks the manifest up to 7 levels deep and picks up any `tools` array or object, so it works on wrapper formats as well as plain listings.
It matches keywords against seven rules: destructive action, code or shell execution, and external write or message are scored high (5 points each); filesystem write, network access and sensitive-data access are medium (3); broad read access is low (1). A tool's overall level is the highest severity it matched, so any single high signal makes it high risk.
No. It only reads the two JSON manifests you paste, and every step — parsing, tool extraction, diffing and scoring — runs in your browser. The generated report states explicitly that no tool was connected or executed.
Description changes, input schema properties added or removed, fields newly listed as required, and any change to the tool's annotations, permissions or capabilities block. Comparison uses a key-sorted serialisation, so pure formatting or key-order differences are ignored.
No — this is keyword-based static analysis of declared metadata, and a manifest can describe a capability in wording none of the seven rules match, or a server can behave differently from what it declares. Treat the output as a review checklist and inspect the server's actual code and network behaviour before granting it access.