MCP Config Guard - agent config lint

It reads an MCP client config - .mcp.json, .vscode/mcp.json, mcp.json or claude_desktop_config.json - and flags the lines that give an AI agent more reach than intended. It checks 23 rules across four groups: how the server code is fetched, which credentials it is handed, how far into the disk it reaches, and how it is reached over the network. Each finding names the line, the risk and the replacement.

Engineers who added MCP servers to a shared repository so Claude Code, Copilot, Cursor or Windsurf agents can use them, and platform or security engineers who now own a config file that several people edit and nobody re-reads. It is also for reviewers who see .mcp.json in a pull request and have no checklist for it.

The existing scanners connect to each server and read its tool descriptions, which means starting the code you are trying to judge, and they run as a separate command outside the editor. A chatbot needs the file pasted into it, credentials and all. This reads the text of the file you already have open, offline, and starts nothing.

Checking the config file you have open is free, forever, with all 23 rules, no watermark and no limit on how many files you check one at a time. The licence key covers scale and hand-off: one sweep over every config in the workspace and the client config folders, a dated CSV, JSON or HTML report, and machine-readable output for CI.

Upwork lists cybersecurity developers at a $60 median hourly rate, $40 to $90 typical, as of September 2026. A person reading every MCP config in a monorepo and writing the findings down is an hour or two each time somebody adds a server. The extension is $29 once, one licence key per person or CI seat, with a 7-day full refund.