pi-mono

Overview

pi-mono is the closest thing in the open-source ecosystem to a “build-your-own coding agent” starter kit. Maintained by Mario Zechner under the badlogic GitHub handle, the monorepo packages every layer a developer needs to ship a working agent product: the agent loop itself (CLI), the model abstraction (unified LLM API), the user-facing surfaces (terminal and web UI libraries), an integration template (Slack bot), and the deployment substrate (vLLM pods).

The repo crossed 39K stars on April 23, 2026 and has trended consistently for over a month — sustained, not viral. Developers are using pi-mono as both a reference implementation (how does a real agent harness work end-to-end?) and a fork point (clone it, replace the agent loop with your domain logic, ship).

Why It Matters

The agent-tooling trending lane in April 2026 has split into two camps. One camp ships configuration layers — cc-switch, mattpocock/skills, awesome-codex-skills — that sit on top of existing tools (Claude Code, Codex). The other camp ships runtime layers — opencode, jcode, pi-mono — that give you the actual agent infrastructure to build new tools. pi-mono is the most complete entry in the runtime camp: it doesn’t just give you the agent loop, it gives you everything around the loop you’d otherwise build yourself.

This matters because the moat for AI coding products is moving from “model access” (commoditized) to “harness quality” (differentiated). Teams that want a custom coding agent — domain-specific, brand-aligned, deployable on their own infrastructure — can get to a working v1 in days using pi-mono instead of months building from scratch.

Key Capabilities

Coding agent CLI: A working agent loop you can run today — pnpm install && pnpm dev and you have a Claude Code-style terminal experience. Replaceable agent logic, swappable providers, configurable tool sets.

Unified LLM API: A single TypeScript interface that abstracts Anthropic, OpenAI, Google Gemini, and local OpenAI-compatible endpoints (Ollama, vLLM, llama.cpp). Switch providers via config, not code rewrites.

TUI and web UI libraries: Reusable components for both terminal and browser surfaces. The TUI library handles streaming output, syntax highlighting, and tool-call rendering; the web UI library ships with chat, file-tree, and diff-view components.

Slack bot harness: Reference implementation for wiring an agent into Slack channels — DMs, threaded replies, ephemeral responses, file uploads.

vLLM pods: Kubernetes manifests for deploying your own model-serving infrastructure with vLLM. Useful for teams running local Llama/Qwen/DeepSeek models.

Getting Started

git clone https://github.com/badlogic/pi-mono
cd pi-mono
pnpm install
pnpm dev

The monorepo’s apps/ directory contains runnable examples; packages/ contains the reusable libraries. Documentation is the README plus inline TypeScript types — the codebase is the spec.

Who It’s For

Developers building custom AI coding tools who want a working harness without the six-week setup tax. Teams shipping domain-specific agents (legal, medical, finance) where model + skill configuration on top of Claude Code or Codex isn’t enough — they need their own runtime. Engineering organizations evaluating whether to build agent tooling internally vs. license third-party platforms; pi-mono is the strongest “build” option in open source.