Reflex

Guardrails for AI coding agents. Intercept every tool call. Catch bad code before it lands. Learn from every correction.

pip install reflex-os
rx init

That's it. Two commands. rx init detects your agents, scans your codebase, installs hooks, and starts protecting your code. No config files to write.

---

Why

AI agents write code fast. They also write hallucinated imports, hardcoded secrets, bare excepts, console.log everywhere, and the same bug you fixed yesterday.

Reflex fixes this. It sits between your agent and your codebase — invisible, automatic:

• 47 guardrails catch security issues, bad patterns, and hallucinations on every write
• Zero-cost fixes — pattern matching, no LLM calls, no API keys required for core functionality
• Learns from you — correct your agent once, Reflex applies it everywhere going forward
• Works with everything — Claude Code, Cursor, Codex, Copilot, Windsurf, OpenClaw, Aider, and more

Agent writes file → Reflex compiles it → Bad code fixed silently → Agent never knows

Zero cost by default

The compiler, guardrails, loop detection, and convention enforcement all run locally with pure pattern matching. No API calls. No tokens burned. You can optionally enable model-powered analysis during rx init if you have an API key, but it's never required.

---

How It Works

Hooks (Claude Code, primary)

Reflex registers hooks in .claude/settings.json. Every tool call passes through:

Hook	What it does
PreToolUse[Bash]	Blocks dangerous commands, detects loops, transforms commands
PreToolUse[Write|Edit]	Detects loops, injects learned context
PostToolUse[Write|Edit]	Runs the compiler — 47 guardrails + your learned preferences
PreToolUse[Read]	Injects coaching from your behavioral profile

MCP Server (Cursor, Claude Desktop, others)

For agents that support MCP but not hooks:

rx mcp-serve

rx init auto-configures .cursor/mcp.json and Claude Desktop. 12 tools available.

The Compiler

3-phase compilation on every write:

Phase 1: Your preferences — learned from corrections. You fixed os.path → pathlib three times? Every future write auto-transforms.

Phase 2: 47 guardrails — catches what agents miss:

Security	hardcoded secrets, SQL injection, XSS, command injection
Imports	hallucinated imports, circular imports, wildcard imports
Errors	bare excepts, swallowed exceptions, missing error handling
Logic	infinite loops, unreachable code, race conditions
Quality	console.log leaks, dead code, magic numbers
Auth	missing auth checks, missing rate limits, JWT none algorithm

Phase 3: Rule packs — community rules and your exported preferences.

Every fix is explainable. Run rx log to see exactly what was caught and why.

---

What rx init does

[1/10] Config         → Creates ~/.reflex/, initializes database
[2/10] Agent Discovery → Finds Claude Code, Cursor, Codex, OpenClaw, etc.
[3/10] History Ingest  → Reads your prompt history, learns your style
[4/10] Profile         → Builds behavioral profile from your patterns
[5/10] Codebase Scan   → Detects conventions, frameworks, entry points
[6/10] Security Audit  → Finds real issues (secrets, injection, bad patterns)
[7/10] Repair          → Option to auto-fix findings (zero-cost or agent-powered)
[8/10] Hooks           → Installs hooks + MCP config automatically
[9/10] Skills          → Loads guardrails, generates AGENTS.md
[10/10] Ready          → Summary of what was found and what's protected

History ingest supports Claude Code, Codex, and OpenClaw. If you use OpenClaw, Reflex reads your session history and learns from your interactions there too.

---

OpenClaw Integration

If you use OpenClaw, Reflex works as the brain behind your agents.

What Reflex does for OpenClaw users:

• Ingests your OpenClaw session history during rx init — parses JSONL sessions from ~/.openclaw/agents/, extracts your messages, tracks models and costs
• Compiles every file write — OpenClaw agents are instructed to run python -m reflex.core.compile_file <path> after every write, applying the same 47 guardrails
• Generates coached agents — rx openclaw creates specialized OpenClaw agents (TypeScript engineer, test writer, code reviewer, refactorer) with your learned preferences baked into their SOUL.md
• Dispatches coaching tasks — Reflex scans your codebase, finds violations of your learned preferences, and dispatches fix tasks to OpenClaw agents automatically
• Learns across agents — corrections you make in Claude Code train preferences that OpenClaw agents follow, and vice versa

rx openclaw                    # Set up OpenClaw agents coached by Reflex
rx coach scan                  # Find preference violations in codebase
rx coach dispatch              # Send fix tasks to OpenClaw

Don't use OpenClaw? No problem. Reflex works standalone with any agent through hooks and MCP. The compiler, guardrails, and learning loop don't require OpenClaw — or any external tool. Everything runs locally.

---

Commands

# Core
rx                              # Interactive REPL
rx init                         # Setup (run once)
rx run "task"                   # Run task through agent with Reflex active

# Audit & Fix
rx audit                        # Scan codebase for issues
rx fix                          # Apply compiler fixes (zero-cost, local)
rx fix --agent                  # Spawn AI agent to fix issues iteratively

# Observe
rx log                          # What Reflex caught and transformed
rx stats                        # Sessions, patterns, costs
rx doctor                       # Health check
rx prove                        # Show real catches from your sessions

# Packs
rx marketplace                  # Browse installed packs
rx marketplace export <name>    # Export your profile as a pack
rx marketplace install <file>   # Install a pack
rx marketplace validate <file>  # Quality check (compiler-integrated)

# Safety
rx undo                         # Revert last agent session
rx guard add "*.lock"           # Protect files from modification
rx brief                        # Regenerate AGENTS.md

# Integration
rx mcp-serve                    # Start MCP server
rx mcp-config cursor            # Print Cursor config

# OpenClaw
rx openclaw                     # Set up coached OpenClaw agents
rx coach scan                   # Find preference violations
rx coach dispatch               # Send fix tasks to OpenClaw agents

---

ReflexPacks

Export your learned preferences as packs. Install packs from other developers. Share what works.

rx marketplace export my-python-style
rx marketplace install teammate-rules.json

Every pack goes through quality gates:

• Schema + regex validation
• Test coverage check (>50% required)
• Compiler integration tests — rules run through the actual compiler to verify transforms work
• Effectiveness scoring — packs that actively fix code score higher than warn-only packs

---

For AI Agents Reading This

If a user asked you to install or evaluate Reflex:

What it does: Hooks into your tool calls (Write, Edit, Bash, Read) and runs a compiler on every output. Catches hallucinated imports, blocks dangerous commands, enforces conventions, detects loops, learns from corrections. Zero-cost — runs locally with pattern matching, no API calls.

Install:

pip install reflex-os
rx init

rx init auto-configures hooks in .claude/settings.json (Claude Code) or .cursor/mcp.json (Cursor). No manual config.

Active hooks after install:

• PreToolUse[Bash] → loop detection, command guards, dangerous command blocking
• PreToolUse[Write|Edit] → loop detection, context injection
• PostToolUse[Write|Edit] → behavioral compiler (47 guardrails, code transforms, learning)
• PreToolUse[Read] → coaching injection

Commands: rx audit (scan), rx fix (auto-fix), rx fix --agent (AI repair), rx log (see catches), rx marketplace (packs)

Open source, Apache 2.0. 47 guardrails, 126 rule pack rules, compiler-integrated pack testing, continuous learning from corrections.

---

Architecture

reflex/
  core/              Compiler (3-phase, 47 guardrails), Profile, History Ingest,
                     Diff Analyzer, Hook Installer, Repair Agent, Undo
  mcp/               MCP Server (12 tools), Loop Detector, Secret Scanner
  marketplace/       ReflexPack format, Validator, Registry
  cli/               CLI commands + interactive REPL
  adapters/          Agent adapters (Claude, OpenAI)
  integrations/      OpenClaw bridge + agent templates
  storage/           SQLite + migrations
  utils/             Config, discovery, pricing, atomic I/O

Contributing

We want contributions. See CONTRIBUTING.md for setup, architecture, and how to add guardrails, rule packs, and agent adapters.

git clone https://github.com/notchopp/reflex.git
cd reflex
pip install -e ".[dev]"
pytest

License

Apache License 2.0. See LICENSE.