MiSeGuard adds deterministic security checks to MCP tools for AI coding agents

Summary

MiSeGuard is an independent Node.js tool that places a stdio proxy between an AI coding agent and its Model Context Protocol (MCP) servers, inspecting tool calls before they reach the operating system. It assigns commands and file operations a deterministic blast-radius score from 0 to 100; the same action under the same configuration always receives the same score, without an LLM in the scoring loop. Green actions are allowed, yellow actions trigger a dry-run in strict mode or a warning in permissive mode, and red actions are blocked in either mode. Users can configure thresholds, command allowlists, and protected paths, with sensitive-file access or changes elevated to the red tier. CLI commands generate agent configuration snippets, wrap existing MCP configurations with a backup, check commands, and simulate them in an ephemeral sandbox; the project says configuration wrapping is idempotent. The README reports policy-evaluation latency below a millisecond in its listed scenarios and points to a benchmark methodology and a command for reproducing the tests. Coverage is limited to MCP traffic over stdio JSON-RPC: native IDE tools that bypass MCP are not intercepted, and prompt injection, HTTP or gRPC transports, and kernel-level bypasses are outside its protection. The project requires Node.js 18 or later.