Coding Tools MCP v0.2.2 Exposes 20 Tools With Landlock Kernel Confinement and Three Permission Tiers
GitHub / Hacker News·medium signal
Coding Tools MCP shipped v0.2.2, a model-neutral MCP runtime that gives 'any AI chat or agent a safe pair of hands on your codebase' via 20 tools across four groups: files/search (read_file, search_text, apply_patch, view_image), execution (exec_command, write_stdin, read_output, kill_session, request_permissions), git (status, diff, log, show, blame), and runtime introspection. The security model is the interesting part for anyone wiring agents into a repo — three permission modes (safe, trusted, dangerous), kernel-level filesystem confinement on Linux via Landlock, one workspace root per server, and absolute path enforcement. The project reports serialized tool-result bytes dropped 37% release-over-release through better summarization, pagination, and result capping. Apache-2.0, 558 stars, 100 forks; works with Claude Desktop, Cursor, and Cline.