Pi

What it is

Pi is an open-source minimal agent harness from Armin Ronacher, who wrote Flask, and Mario Zechner, who wrote libGDX. It lives at earendil-works/pi on GitHub and passed 50,000 stars within months of launch — unusually fast for a project whose pitch is that it does less than the alternatives.

The pitch is exactly that, though. Pi's system prompt is under a thousand tokens. Most harnesses ship several thousand words of instructions, examples and scaffolding designed to coax useful behaviour out of a model. Pi's argument is that current frontier models do not need the coaxing, and that the scaffolding is now getting in the way — burning context, imposing a workflow, and papering over capability the model already has.

How the loop works

A terminal agent with a small tool set and almost no opinion about how you should work. You bring a model; Pi passes your request through with minimal framing and lets the model decide how to proceed. Where a heavier harness would insert a planning phase, a summarisation step or a prescribed edit format, Pi generally does not.

The practical consequence is that Pi's quality tracks your model's quality much more directly than usual. On a strong model it feels clean and fast, because nothing is competing with the model's own judgement. On a weaker one there is no scaffolding to catch it, and the results reflect that.

Isolation and permissions

Tools are permissioned — the agent asks before it acts, in the pattern established by the mainstream terminal agents. There is no sandbox of its own, so containment is whatever you put around the process. Given how little the harness interposes, the permission prompt is doing more of the safety work here than in tools with additional guard rails, and it is worth reading the prompts rather than approving on reflex.

Who it is for

People paying for a frontier model who would rather not have a harness spend their context window telling that model how to be a coding agent. It has become a regular fixture in the community's hard-mode harness comparisons alongside Claude Code and Codex, which is the clearest signal that the minimal approach holds up under real workloads rather than only in demos.

It also suits anyone who wants to read the whole harness before trusting it. A small codebase with a short prompt is auditable in an afternoon.

Limits

Two things we have not verified for this entry and will not guess at: the precise licence class and the implementation language. The project is open source, but check the repository for the specific licence before you make a compliance decision on it.

Beyond that, the minimalism cuts both ways. There is less here — fewer integrations, less workflow tooling, less handling of the awkward cases a mature harness has accumulated fixes for. On a weak or cheap model you will feel the absence of scaffolding. And the project is young, so expect churn.

Alternatives on this board

• Claude Code — the heavyweight comparison, with permissions and tooling built out.
• Codex CLI — open source and sandboxed by default.
• OpenCode — provider-agnostic, similar spirit, more batteries included.

Sources

• earendil-works/pi on GitHub — licence and implementation details are stated there