Pi vs Claude Code: setup and extensions

Pi is a coding assistant with a small core that users can extend. This video compares its approach with Claude Code and Codex and explains how a harness controls a model's tools, context and execution loop. The presenter discusses the SWE-agent harness-ablation research and benchmark examples to argue that tool design can affect results even when the model stays the same.

The Pi walkthrough covers file tools, instructions through AGENTS.md and sessions stored as branchable JSONL logs. Its provider configuration accepts OpenAI-compatible endpoints, including a local LLM running on the user's machine. The presenter describes this flexibility as a reason to consider Pi for custom workflows.

Setup covers installation through npm, account login or an API key, and a repository summary inside a Docker container. The presenter stresses that Pi's default setup lacks permission prompts and built-in safeguards, so users must arrange containment themselves. In the version described, the presenter says MCP support, sub-agents and plan mode are absent from the core.

The demonstration adds web access and sub-agent extensions from the package catalog. Community extensions can conflict, and users can instead write TypeScript extensions to register tools or gate tool calls. The presenter recommends Claude Code or Codex for most less technical users, while positioning Pi for people willing to manage customization and security. The closing advice is to test each model and harness combination on actual work.