Summary

Prax is a self-improving agent runtime for LLM agents, designed to learn from experience and execute tasks reliably within code repositories. It targets developers and power users needing robust automation, driving agents through test-verify-fix loops to enhance accuracy and reduce errors in development workflows.

How It Works

Prax employs a verification-first architecture centered on a test-verify-fix loop: it executes tests, analyzes failures, edits code, and re-runs until success. This is augmented by experience-based self-improvement, learning from corrections and task completions across sessions via persistent JSON/SQLite memory. It also supports multi-model orchestration for diverse LLM providers.

Quick Start & Requirements

Requires Node.js (v14+) and Python 3.10+. Install CLI via npm install -g praxagent. Configure LLM services using LLM_BASE_URL/LLM_API_KEY env vars and ~/.prax/models.yaml. Claude Code IDE integration is available via prax install --profile full. Native Windows support is on the 0.5.x roadmap.

Highlighted Details

• Achieves 10/10 success on repository repair tasks, 49% faster than baseline.
• Core verification-first architecture with test-verify-fix loops.
• Cross-session persistent memory (JSON/SQLite) and global experience store.
• Multi-model orchestration supporting various LLM protocols.
• Security features include permission modes (read-only, workspace-write, danger-full-access).
• Includes 25+ tools and bundled skills, with commercial recipes for team workflows.

Maintenance & Community

The provided README does not detail maintainers, sponsorships, or community channels.

Licensing & Compatibility

Released under the MIT License, permitting broad commercial use.

Limitations & Caveats

Native Windows support is planned for the 0.5.x roadmap. Some bundled commercial recipes have built-in restrictions to prevent destructive actions, requiring careful review for specific enterprise workflows.