Pi Coding Agent - The coding-agent harness you can make your own

Hi everyone!

Pi is a minimal, hackable terminal harness for building the AI coding agent workflow you actually want.

It keeps the core small and clean, then gives you the freedom to add, modify, or replace almost anything through extensions, skills, and packages. Context management, sub-agents, permissions, custom workflows — you can make it behave exactly how you like.

Pi first caught a lot of attention as the harness underneath @OpenClaw. What’s really striking now is how strongly it’s showing up in the @OpenAI ecosystem. It’s astonishing that Pi and @opencode have roughly similar usage share in OpenAI/Codex production traffic!

It’s cool to own your harness 🤓

Pi is such a joy to use! It's easy, customizable, and fits my workflow perfectly. I use Pi inside my VScodium terminal with opensource models. Pi makes my entire dev workflow FOSS, which means I no longer worry about large tech companies making changes to their models.

We’re entering the era where coding agents become teammates instead of tools. Interesting direction.

Even though my ecosystem is more of shifted to KiloCode ocasionally and VSCode for the rest of self coding or building tasks, Pi agent does seem very useful for such a simple agent that you get to decide what it can or can't do and build from a simple base.

Also the way it is handling chat forks is much better than other bots i have seen, Copilot is not that great while kilocode was better than copilot but this is the style i'd much prefer, asking questions to clarify in middle of chat and then turn back and continuing like we never did in first place, but made this interaction much more seemless

@Pi Coding Agent The customizable harness angle is interesting. I think coding agents get much more useful when teams can shape the workflow around their own review habits, instead of adapting everything to one fixed agent style.

The npm-based package distribution is the right architectural call. Making the agent extensible through tooling developers already use for dependency management avoids the 'special plugin registry' trap. We've wasted time fighting opinionated defaults in other coding agents, so shipping without sub-agents and plan mode by design is a real differentiator. How does prompt template inheritance work across packages? Can a base template be partially extended?

I'm using it as my daily driver for local LLMs and it's amazing! There's one thing I'd really want to see there: the ability to see the diffs for all the files that were changed in a given session. The list of files is already there with additions/deletions stats, so I assume that the content should be available as well (therefore can be displayed)

I like the direction @Pi Coding Agenttakes in making the harness small and extensible rather than hiding everything behind a fixed workflow. Coding agents become really useful in serious codebases when teams can see the context boundary, control permissions, customise the workflow, and still review the diff like normal engineering work.