Introduction
agentreflex is the open commons of reflexes for AI coding agents — write an instinct once, and it fires in every agent.
AI coding agents edit files, run commands, and call tools on your behalf. agentreflex lets you capture the instincts you'd apply by hand — reflexes — and have them fire in every agent you use: Claude Code, Cursor, Gemini CLI, Copilot CLI, Windsurf, and OpenCode natively, advisory everywhere else.
Skills vs. reflexes
| Layer | The question it answers | Where it lives |
|---|---|---|
| Instructions | What should the agent know and prefer? | AGENTS.md |
| Skills / tools | What can the agent reach and do? | MCP, skills |
| Reflexes | What would you do the moment it acts? | agentreflex |
Instructions are advice a model can ignore. Skills hand it new powers. Neither guarantees anything at the instant an action runs. A reflex does — deterministic code that sees every tool call and decides, before it happens.
What a reflex is
A small module that reacts to what an agent is about to do:
export default {
name: "no-force-push",
onToolCall(ctx) {
if (ctx.tool === "Bash" && /git\s+push\b.*--force\b/.test(ctx.command ?? ""))
return { action: "deny", reason: "no force-push — open a PR instead" };
return { action: "pass" };
},
};Write it once. agentreflex compiles it into each agent's native hook, so the same instinct fires everywhere.
An open commons
A reflex is a file you own — keep it in your repo, hand it to your team, or pull one the community built:
npx agentreflex add no-force-push