Pi 1.0: Open Source Coding Agent Goes Stable
Earendil released Pi 1.0, an MIT-licensed terminal coding agent. What the stable release adds, and when a bring-your-own-model agent pays off.
Earendil released Pi 1.0 on 1 October 2026, the first stable version of its terminal coding agent. The release is MIT licensed, installs with a single shell command on Unix and Windows, and works with models from most providers or with one running on your own machine. Earendil says hundreds of thousands of people use Pi every week.
The interesting part of a 1.0 is not the feature list. It is the promise that the interface will stop moving. For a tool you wire into scripts and CI, that is the whole value.
What the stable release adds
The announcement names a specific set of changes:
Codemode, native support for the Model Context Protocol and for non-LLM models, including the small decision models that have started appearing this autumn.
Extension support for virtual models, so an extension can route each request to a different model.
Deferred tool loading, where tool definitions are fetched when needed rather than pushed into every request.
Cache warming for Anthropic models.
Mid-conversation system messages.
A new terminal theme, with full-screen mode as the default.
Alongside it, Earendil shipped Pi Durable, an experimental package for long-running agentic work that outlives a terminal session. It is explicitly labelled experimental, so treat it as a preview rather than something to put under a production job.
Why deferred tool loading and cache warming matter to your bill
These two sound like plumbing. They are the two items on the list that change what a session costs.
Every tool definition you give an agent is tokens, and those tokens are re-sent on every turn. An agent connected to six MCP servers can spend a large slice of each request restating tools it will never call on that turn. Deferred loading means the definitions arrive when the model asks for them. Fewer tools in the window also tends to mean better tool selection, which is the quieter benefit.
Cache warming attacks the same problem from the other side. Prompt caching lets a provider charge less for input it has already processed, and warming keeps that cache alive across the gaps in a working session rather than letting it expire between your coffee and your next instruction. If you want the mechanics, we have a separate explainer on how prompt caching changes what a long session costs.
Where a bring-your-own-model agent fits
The real decision Pi 1.0 puts in front of you is not Pi versus another agent. It is whether you want the agent and the model bought together or separately.
A subscription agent bundles both. You pay one price, the vendor picks the model, and when they change it your results change with it. A bring-your-own-model agent like Pi splits them: the harness is free and auditable, and you pay the model provider directly at API rates. That is usually cheaper at low volume and more expensive at high volume, which is the opposite of what people expect. Working out which side of that line you are on is the exercise in what an AI coding agent costs per month.
The MIT license matters for a second reason. If the project stalls, the harness you already installed keeps working, and you can fork it. That is a weak form of insurance, but it is more than a closed agent offers. For how these harnesses are put together, see what an agent harness actually is.
What to check before switching
Pi being stable does not make it the right agent for you. Three things decide that:
Which models you already pay for. A harness that routes to your existing provider keys avoids a second bill. One that needs a new subscription does not.
Whether you need a GUI. Pi is a terminal tool. If your workflow lives in an editor, a terminal agent is an addition rather than a replacement.
How much of your tooling is MCP. Codemode is most useful if you already have MCP servers worth connecting. If you have none, it is a feature you will not touch this quarter.
If you are weighing it against what you run now, the honest test is to run both on the same tasks from your own repository rather than on a benchmark. Benchmarking coding agents on your own codebase covers how to set that up without fooling yourself.
The wider pattern
Pi 1.0 landed on a day thick with agent releases, and the pattern across them is worth noting: the harness is increasingly free, and the money is in the model. That is good for anyone building, because it moves the lock-in from the tool you type into to the API you call, and API choices are easier to reverse than habits.
Sources: Earendil's Pi 1.0 announcement and the Pi documentation.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
