Pi Coding Agent

About

Pi Coding Agent is the open-source coding CLI that actually does what every vendor pretends to do: it runs Claude, GPT-5, Gemini, Grok, and your local Llama in one harness, with no CLI seat fee. Mario Zechner built it. 56,300+ GitHub stars at time of writing. MIT licensed. Version 0.76.0 shipped on May 27, 2026 — one day before this listing. What Pi Actually Is Pi (github.com/earendil-works/pi) is a minimal terminal harness for an agent loop. The primitives are deliberately small: a planner, an executor, a tool layer, and a unified provider-agnostic LLM API that abstracts every major model behind a single interface. You bring the API keys. You pick the model per session — or per turn, inside the same conversation. The agent loop runs the same way regardless of who is on the other end. The pitch is direct: 'The coding-agent harness you can make your own.' Pi adapts to your workflow rather than the other way around. Customization is via extensions, skills, prompt templates, and themes that ship as Pi packages and are distributed via npm or git. The community catalog is growing weekly. The Monorepo Is The Whole Story The Pi monorepo bundles five projects against the same primitives. The coding agent CLI is the headline. The unified LLM API is the abstraction layer that makes everything else interchangeable. Terminal and web UI libraries cover both surface types. A Slack bot harness lets you drop the same agent loop into a team chat channel. vLLM deployment pods package the production inference pipeline. The reason this matters is that it removes the rewrite tax. You prototype on your laptop with Claude. You decide to move to Gemini for cost. You ship to production behind a Slack bot with Llama running on your own vLLM pods. The primitives do not change. The configuration does. That is the difference between OSS infrastructure and a vendor toolchain. The Money Math Most working developers in 2026 are running three or more AI coding assistants in parallel — Claude Code for refactors, Cursor for IDE work, Codex CLI for OpenAI-specific tasks. Each one is $20/month. The combined seat spend is $40-60/month per developer before any tokens are consumed. For a 10-person engineering team, that is $400-600/month in CLI subscriptions alone, on top of the model API spend. Pi consolidates the harness. You still pay the model providers for tokens consumed, but the CLI itself is free under MIT. For a 10-person team running mixed-model workloads, that is roughly $5,000-7,000/year in pure CLI seat fees that disappear from the budget. The math is brutal when you ladder it up to the org level. Who Should Install It This Week Developers running 3+ AI assistants in parallel should install it immediately — the consolidation pays back inside a week. Teams hosting local models on private inference (vLLM, llama.cpp, Ollama) should install it because Pi treats local models as first-class backends, not afterthoughts. Engineering leads who refuse single-vendor lock-in on AI tooling should install it as their default CLI and demote the vendor-specific ones to model-specific roles. If you are deep inside Claude Code and dependent on Anthropic-specific features like Computer Use, Pi is not a drop-in replacement yet. It is a complement, not a swap. The right play is to install Pi for cross-model experimentation and keep Claude Code for Anthropic-native workflows. The Bigger Pattern Pi is part of the quietly-growing 2026 pattern of open-source AI infrastructure that hands authority back to the developer. CodeGraph pre-indexes your codebase as a semantic graph so the agent reads only what matters — Pi pairs naturally with it. Code Review Graph MCP drives a 38x-528x token reduction on code review by feeding the agent only the blast radius of a change. The pattern across the stack is the same — constrain what the AI is allowed to do, hand the human the steering wheel, and watch the failure rate collapse. The AI-replacement-narrative debunk we published this week walks through the data on why this is the right architecture.

Key Features

• Provider-agnostic by design: one CLI runs the same agent loop against Claude, GPT-5, Gemini, Grok, or any local model you can host (Llama, Qwen, DeepSeek, Mistral).
• Model switching is supported within a single conversation session — drop GPT-5 for a planning step, swap to Claude for the implementation, finish on Gemini for cheap iteration.
• Bring your own key. No CLI seat fee. You pay the model provider directly for tokens consumed (or zero if the model runs locally).
• Monorepo bundles a coding agent CLI, unified provider-agnostic LLM API, terminal and web UI libraries, a Slack bot harness, and vLLM deployment pods — same codebase covers laptop to production.
• Extensions, skills, prompt templates, and themes ship as Pi packages and are distributed via npm or git. The community catalog is growing weekly.
• Ships with powerful defaults but deliberately skips opinionated features like sub-agents and plan mode — install only what you want.
• MIT license. 56,300+ GitHub stars at time of listing. v0.76.0 released May 27, 2026.

Use Cases

• 1Developers who built their workflow inside Claude Code but want to test GPT-5 or Gemini without rewriting their agent setup.
• 2Teams running local models (Llama 4, Qwen 3, DeepSeek) on a private inference endpoint and need a polished CLI that does not assume an OpenAI or Anthropic key.
• 3Solo developers running 3+ AI assistants in parallel who want to consolidate $40-60/month in CLI seat fees into a single open-source harness.
• 4Engineering leads who refuse single-vendor lock-in on AI infrastructure and want a tool that survives an Anthropic price hike or an OpenAI policy change.
• 5Indie hackers and OSS maintainers who want to publish their own extensions and skills as Pi packages without going through a closed marketplace.
• 6Teams running a Slack bot harness on top of the same agent loop they use locally — the monorepo covers both surfaces.