rex-machine 0.3.0

Extract lessons learned (REX) from code repositories

rex-machine

Your codebase has lessons no one wrote down. This finds them.

rex-machine scans any code repository and extracts structured technical lessons learned — what problems the team faced, what they tried, what worked. Claude navigates your repo autonomously (reading files, grepping, following imports) and produces a grounded report, not a vague summary.

Works on any language, any framework, any repo size.

PyPI · GitHub

Install

pip install rex-machine

Usage

# Set your API key (one time)
export ANTHROPIC_API_KEY=sk-ant-...

# Analyze a repo
rex extract /path/to/repo

# Save as Markdown
rex extract /path/to/repo -o markdown -f report.md

# Save as JSON
rex extract /path/to/repo -o json -f report.json

# Analyze in French
rex extract /path/to/repo --lang fr

# See what Claude is reading in real time
rex extract /path/to/repo -v

What you get

Each finding is structured, not prose:

REX #1: Error Handling Strategy                    [Strongly Inferred]

  Context:    REST API with 23 endpoints
  Problem:    Inconsistent error responses across modules

  Approach             Worked?   Details
  Global exc handler   Yes       Centralized in middleware
  Per-endpoint catch   No        Led to code duplication

  Learning:        Centralized error handling > scattered try/catch
  Recommendation:  Implement global exception middleware early
  Source files:    src/middleware/errors.py, src/api/deps.py

Every REX has: a problem, approaches tried (what worked and what didn't), a learning, a confidence level grounded in code evidence, and the exact source files.

Multi-provider

Works with any Claude provider. Auto-detects from environment variables.

Provider	Auth
Anthropic (default)	ANTHROPIC_API_KEY
Azure AI Foundry	ANTHROPIC_FOUNDRY_RESOURCE + ANTHROPIC_FOUNDRY_API_KEY
AWS Bedrock	AWS credentials
Google Vertex AI	GCP credentials

rex configure  # interactive setup

How it works

Claude gets three read-only tools (read_file, list_files, grep) and decides what to explore on its own — same agentic loop as Claude Code. No files are pre-selected by heuristics. Claude reads what it needs, follows leads, and stops when it has enough evidence.

All tools are read-only and sandboxed to the repository. No writes, no network, no code execution.

Built with the raw Anthropic SDK. No LangChain, no agent framework. The core loop is a while stop_reason == "tool_use" — about 20 lines.

Configuration

rex init          # create .rex-machine.json in your project
rex configure     # set up provider & credentials
rex status        # show current config

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

License

Apache 2.0

Star History