MCP gives your agent hands. agentreflex gives it reflexes.

Coding agents edit files, run commands, and call tools on your behalf — each through its own hook system, none of which talk to each other. agentreflex is the reflex layer that sits under all of them: write an instinct once, and it fires natively in every agent you use.

A reflex is the thing you'd do by hand the moment an agent acts — reach for the test, flinch at a force-push, refuse to read a .env, pause before prod. Not advice the model can ignore. Deterministic code that sees every tool call and decides, before it happens.

import { defineReflex, deny, pass } from "@agentreflex/core";

export default defineReflex({
  name: "no-force-push",
  onToolCall(ctx) {
    if (ctx.tool === "Bash" && /git\s+push\b.*--force\b/.test(ctx.command ?? ""))
      return deny("no force-push — open a PR instead");
    return pass();
  },
});

agentreflex compiles it into each agent's native hook, so the same instinct fires across Claude Code · Cursor · Copilot CLI · Gemini CLI · Windsurf · OpenCode — and as a compiled AGENTS.md advisory everywhere else.

npx agentreflex init

One command wires every agent in your repo to your .reflex/ folder. After that, editing a reflex is instantly live — no rebuild, no re-publish.

no-force-push · no-rm-rf · no-secrets · ask-on-prod · stay-in-repo · conventional-commits · recover (snapshots files before edits) · abide (your human↔agent working agreement, enforced) — browse the catalog →

Reflexes run in your dev loop, so the runtime is built to be trusted: agentreflex add shows the code and the reflex's declared capabilities before anything is wired in, and the dispatcher fails open — it never blocks or breaks your agent because of its own error.