Pi Coding Agent is an open-source terminal coding agent that reads, edits and runs code in your project folder using the model provider you choose. Extensions, skills and packages add the features it leaves out by design.
Used this tool? Rate it
Used this tool? Rate it
Pi Coding Agent (called simply "Pi" on its own site) is a terminal coding agent: you give it a goal and a working folder, and it can inspect files, run commands, edit content and work through multi-step tasks.
The project positions Pi as a minimal agent harness that you adapt to your workflows rather than the other way around. Instead of shipping every feature in the core, it keeps a small base and expects you to add what you need through extensions, skills, prompt templates, themes and packages.
Pi was written by Mario Zechner, and Earendil Inc. announced on April 8, 2026 that it had acquired the Pi open-source project. The pi.dev site footer now credits "Earendil Inc. & Contributors" and links to the MIT License.
Before choosing it, know three things: Pi has no price of its own but needs a model provider, it acts without asking for per-step approval, and features such as sub-agents, plan mode and MCP support are left out of the core on purpose.
Pi's core is a small tool set, a session store and a loader for context files. Almost everything else is a customization layer on top.
With bash and write enabled by default, an unmodified session can change files and run shell commands; narrowing the tool list is the built-in way to reduce that reach.
The homepage lists the kinds of features extensions are expected to supply instead of the core: sub-agents, plan mode, permission gates, path protection, SSH execution, sandboxing, MCP integration, custom editors, status bars and overlays. It also describes Pi as able to modify itself: if you need a command, tool, provider, workflow or UI tweak, you ask Pi to build it, then reload and keep working. A public package catalog on pi.dev lists community extensions and skills published to npm, so a missing capability may already exist as a package.
The quickstart gets a first session running in five steps:
Pi fits people who want to shape their own agent more than they want a finished product.
It is a weaker fit for anyone who expects per-action approval prompts out of the box. The May 2026 how-to noted that Pi's speed comes at the expense of many guardrails and safety features seen in other coding agents and suggested running it in a virtual machine, container or even a Raspberry Pi.
Pi needs access to a model through a supported provider, which can be a subscription, an API key or a local model. No price, plan or checkout for Pi itself appeared in the pages checked for this review, and pi.dev had no separate pricing page. The MIT License it ships under grants permission to use, copy and modify the software free of charge.
The models guide pairs each kind of model access with a setup path in Pi:
| What you have | Setup in Pi |
|---|---|
| A supported subscription | Sign in through /login |
| A provider API key | Store it through /login or set the provider's environment variable |
| A local GGUF model | Connect Pi to the llama.cpp router |
| An OpenAI-, Anthropic- or Google-compatible endpoint | Add it to models.json |
| A provider with a custom protocol or login flow | Build or install a provider extension |
Your spend therefore depends on the provider and model, not on Pi. The interactive footer shows context usage plus accumulated usage and cost, and the model catalog on pi.dev lists providers, model IDs, context limits and pricing. As captured on September 28, 2026, that catalog listed 1,516 models with input and output prices per million tokens.
For long-term planning, Earendil's licensing statement says Pi remains MIT licensed and that this will not change, but that future parts of Pi, or products related to it, may be Fair Source or proprietary. Earendil also runs its own gateway, Radius, which Pi can use as a provider through /login radius; Radius is described as early alpha, and its landing page did not show prices.
Editorial roundups of terminal agents usually compare Pi with tools that ship more features built in. Treat Pi as a Claude Code alternative when control over the harness matters more than out-of-the-box guardrails; the options below differ mainly on that axis.
The July 2026 roundup put the trade-off plainly: Pi can be better than Claude Code for developers who want more control over their coding agent.
Independent coverage points the same way. A May 2026 technology-media how-to described Pi as operating in "YOLO mode" out of the box, with free rein to read and modify anything it has access to.
The homepage lists what Pi deliberately does not include: no MCP, no sub-agents, no permission popups, no plan mode, no built-in to-dos and no background bash. Each comes with a workaround, such as CLI tools documented as skills, a container, a Markdown to-do file or tmux, but none of them is a one-click setting.
Releases can include breaking changes for people who build on Pi. Version 0.87.0, released on September 21, 2026, listed breaking changes including removal of the inherited shouldStopAfterTurn agent option, and 0.87.1 followed on September 22, 2026.
In the pages checked for this review, pi.dev did not publish a separate privacy policy or terms of service.
Pi itself showed no price or plan, and its core is MIT licensed. What you pay depends on the model provider or subscription you connect. Earendil has said that some future additions around Pi may be Fair Source or proprietary.
No. A supported subscription through /login, a provider API key, a local GGUF model through the llama.cpp router, or a compatible endpoint such as Ollama all work.
Not in the core. The project suggests CLI tools documented as skills, or an extension that adds MCP support, and packages from the catalog can provide it.
Anonymous install and update reporting is on by default and can be turned off with the enableInstallTelemetry setting or the PI_TELEMETRY variable. Broader analytics sharing is off unless you opt in.
