Pi Coding Agent
by independent / open source
A minimal, highly customizable coding agent harness built for engineers who want full control over their agentic workflows.
Features
- Minimal by design — 200-token system prompt, extensible from a clean baseline
- Multi-provider support: Codex, Claude, Kimi, MiniMax, Qwen via OpenRouter, GitHub Copilot
- Configuration via
models.jsonandauth.jsonfor provider credentials - Extension marketplace with installable packages (web access, agent extension packs, custom tools)
- Quality-of-life extensions: status lines, permission gates, slow mode, notifications, stashing, question prompts
- Multi-agent orchestration support: one orchestrator, multiple team leads, specialized workers
- Peer-to-peer (Pi-to-Pi) two-way agent communication — agents can send prompts and responses directly to each other
- Theming, widgets, and custom UI via extension system
- Playwright browser agent integration through extensions
- YOLO permissions mode for fully autonomous operation
Superpowers
Pi is the serious alternative to Claude Code for engineers who want to own their agentic harness. Where Claude Code is polished and opinionated, Pi is minimal and fully customizable — you shape the workflow, choose the models, and control the execution environment. The most distinctive capability is peer-to-peer agent orchestration: instead of a strict orchestrator-to-worker hierarchy, Pi agents can communicate directly with each other, preserving context and enabling real collaborative reasoning across devices and model types. This makes Pi especially useful for harness engineering — building reusable, model-agnostic agentic workflows rather than adapting to someone else’s defaults. Practitioners recommend an 80/20 split: Claude Code for mainstream work, Pi for experimental and highly controlled agent workflows.
Pricing
Open source — free to use. Model costs depend on provider.