Harden - a local guard that checks what your coding agent is about to do
Coding agents now run long stretches with nobody watching, and the dangerous command is exactly the one that happens while you are away. Harden runs its own cybersecurity models on your machine and judges every tool call before it executes.
- On their own live counter: 34,222 tool calls checked, 33,382 allowed through, 584 changed or stopped
- Of those, 415 blocked outright and 169 rewritten into a safe version before running
- Works with Codex, Claude Code, Cursor, Antigravity and Kiro; install is one curl line and no account
Judgment happens on the device, so nothing about the repo leaves it - worth an evening if agents already run in your codebase without a human in the chair.
https://harden.run/
📎 Read also:
→ AgentX - test suites and observability around agents
→ A kanban board that dispatches agents with a hard cost cap
→ HackerAI - security audits without the consultant
Coding agents now run long stretches with nobody watching, and the dangerous command is exactly the one that happens while you are away. Harden runs its own cybersecurity models on your machine and judges every tool call before it executes.
- On their own live counter: 34,222 tool calls checked, 33,382 allowed through, 584 changed or stopped
- Of those, 415 blocked outright and 169 rewritten into a safe version before running
- Works with Codex, Claude Code, Cursor, Antigravity and Kiro; install is one curl line and no account
Judgment happens on the device, so nothing about the repo leaves it - worth an evening if agents already run in your codebase without a human in the chair.
https://harden.run/
📎 Read also:
→ AgentX - test suites and observability around agents
→ A kanban board that dispatches agents with a hard cost cap
→ HackerAI - security audits without the consultant