One Daemon. Every AI Agent.
Real-Time Security at the Edge.
CodeMarine hooks into Cursor, Claude Code, Windsurf, Kiro, GitHub Copilot and more through a single daemon architecture. Deterministic scanning in milliseconds on every AI action.
Install the menu bar app and sign in with a verified email. Guard decides on your machine.
What are hooks?
Hooks let security tools intercept AI agent actions before code hits disk. When an AI coding agent writes a file, runs a shell command or calls an MCP tool, the editor fires a hook event - giving CodeMarine a chance to scan the content and block unsafe changes in real time.
PreToolUse
Fires before the agent executes a tool call. CodeMarine scans the proposed action and can block it before anything changes.
PostToolUse
Fires after a tool executes. CodeMarine validates the output and flags anything the agent introduced - the agent can self-correct.
File Write Interception
Catches file modifications before or after they hit disk. Every line the AI writes gets scanned against deterministic rules.
Platform support
One daemon. Six platforms. Every hook event covered.
| Platform | Hook Events | Can Block | Handler Types | Status |
|---|---|---|---|---|
| Cursor | 7 events | Yes | Shell | Beta |
| Claude Code | 18 events | Yes | Shell, HTTP, Prompt, Agent | GA |
| Windsurf | ~5 events | Yes | Shell | GA |
| Kiro | 7+ events | Yes | Shell | GA |
| GitHub Copilot | 4-6 events | Yes | Shell | Preview |
| Antigravity | MCP + Skills | Via MCP | N/A | Coming |
Cursor events include beforeShellExecution, afterFileEdit, beforeMCPExecution and more. Claude Code supports PreToolUse, PostToolUse, Stop and 15 other events.
Why deterministic + hooks is the right architecture
Local daemon. In-memory patterns. Zero cloud dependency. This is what hook-based security should look like.
<10ms daemon calls
All scanning runs on the local daemon. HTTP calls take on your machine - compared to 200-500ms for spawning a new process on every hook event.
No cold start
The full pattern database is loaded in memory when the daemon starts. Every scan hits a warm cache - no initialization overhead on any hook call.
Tier-2 patterns only
Hooks use only high-confidence critical security patterns. Zero noise. No advisory warnings cluttering the agent's context window.
Pre-write scanning
Catches vulnerabilities before code hits disk. The agent gets feedback and can self-correct - you never build on top of a mistake.
Works everywhere
Guard decides on your machine. For Core, your source is never uploaded to scan. Sign-in uses a verified email.
Install once
The same daemon serves all editors simultaneously. Cursor, Claude Code, Windsurf, Kiro - one process protects every AI tool on your machine.
How it works
Daemon starts
CodeMarine daemon runs locally. Pattern database loads into memory. Ready to scan in milliseconds.
codemarine daemon start Hook fires
Your AI editor calls the daemon via hook on every file write, tool call or shell command. One HTTP request per event.
POST local daemon /scan Instant findings
Guard blocks or holds the action before it runs. The decision is made on your machine.
{ "blocked": true, "reason": "..." } The alternative is expensive and slow
AI-based code review tools charge per PR, take minutes and scale with your code volume. Hook-based deterministic scanning does not.
- ⚠ $15-25 per pull request
- ⚠ 20+ minute latency per review
- ⚠ Cost scales with code volume
- ⚠ Non-deterministic - different results on different runs
- ⚠ Code sent to cloud for analysis
- ⚠ Catches issues at PR time - after you've built on the mistake
- ✓ Flat $5/mo - unlimited scans
- ✓ Local per scan
- ✓ Cost scales with nothing
- ✓ Deterministic - same input, same output, every time
- ✓ on your machine - source is never uploaded to scan
- ✓ Catches issues before code hits disk - agent self-corrects
Secure every AI agent on your machine
A menu bar app plus hooks into the agents you use. Coverage varies by agent.
Works with Cursor, Claude Code, Windsurf, Kiro, GitHub Copilot and more - today.