CC-Monitor

English | 简体中文

Ever let an AI agent write code for you and had no idea what it actually did to your machine along the way? Give this a try. It monitors what Claude Code does on your machine — file reads/writes, shell command execution, network access — and blocks or asks for confirmation on high-risk operations, so an AI coding agent can't quietly damage your system or leak data. Everything is audit-logged. The technical design doc is available in English and Chinese. For what risks installing this actually carries, what third-party modules it depends on, and where your data actually goes, see SECURITY.md (Chinese).

Every release — what was added, changed and fixed — is recorded in CHANGELOG.md (Chinese); the current release is v2.0.

🧪 Experimental features (dev branch): master monitors Claude Code only. The dev branch is extending the detection surface to other AI coding agents — Codex CLI, Gemini CLI, Cursor, OpenCode, ZCode, Antigravity CLI and Grok CLI plug into the same rules, approval desk and audit log through their own hooks / plugin; the system-layer eBPF probe recognises every agent's process tree via an "agent registry" and gains file-level syscalls (write / delete / rename / mkdir), listening-port observation, directory-fd tracking (relative paths from rm -r / mkdir -p resolved to absolute ones), session attribution for kernel events and explicit CC-Monitor run -- binding; AI Tap parses Claude Code / Antigravity CLI / Codex session files. Everything except Claude Code is in an experimental / testing stage (implemented from each vendor's docs or source; Antigravity CLI has had one real-machine round, the rest have not been verified on real installs) — not recommended for production. See the dev branch README, MULTI-AGENT.md (usage and internals) and DESIGN-multi-agent.md (design), both in Chinese. To try it: git checkout dev, or git worktree add ../CC-Monitor-dev dev and run it with a separate CC_MONITOR_HOME and port so the master install is untouched.

Core capabilities at a glance