Pi
earendil-works/piReviewed by AI
What it does
Pi is a coding agent you run in a terminal inside a project folder: you describe a task and it reads files, edits them, runs commands and reports back. It also comes as a toolkit of TypeScript libraries (one API for OpenAI, Anthropic, Google and other models, an agent loop, a terminal interface) for people building their own agents, and it can be extended with skills, themes and packages shared over npm or git. It needs Node.js 22.19+ and an AI provider subscription or API key, connected with `/login` inside the program.
How easy it is to install
B · installer plus one simple step · Windows
- Windows B
- macOS B
- Linux B
One install line on each system — `curl -fsSL https://pi.dev/install.sh | sh` on macOS and Linux, `powershell -c "irm https://pi.dev/install.ps1 | iex"` on Windows — and the installers can also put Node.js 22.19+ in place; ready binaries (30 to 43 MB) are attached to each release as well. What keeps it off the top grade is the extra step of signing in to a paid AI provider with `/login`, and that everything afterwards is typed in a terminal.
Numbers on GitHub
- Stars
- 119,188on Oct 31, 2026
- Downloads
- 998,02487,246 this week
- Growth
- +1,153stars in 5 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Node.js
- To build from source
- Node.js
- Disk space
- 43 MB–129 MB · from the installer
- Accounts or keys
- An AI model provider connected with /login — the project's own library lists OpenAI, Anthropic, Google and others
- Cost
- Free, API costs extra · The program is free and MIT-licensed, but it does nothing until you connect a paid AI provider — the README says to run `/login` "to connect a subscription or API key". The cost then depends on the model and how much you use it.
- License
- MIT (Permissive)
- Can replace
- Cursor, Claude Code, Copilot
- Last update
- Nov 1, 2026
Risk signals (6)
- AttentionInstall command runs a script from the internet
Install command pipes a remote script into a shell: curl -fsSL https://pi.dev/install.sh | sh
automatic check
- Attentionno permission system the agent has your full rights
README, Permissions & Containerization: "Pi does not include a built-in permission system for restricting filesystem, process, network, or credential access. By default, it runs with the permissions of the user and process that launched it." It suggests you containerize or sandbox it yourself. So anything the model decides to do — change files, run shell commands, reach the network — happens with your own access, and the same goes for extensions and "Pi packages" shared over npm or git.
AI review
- Infowindows install also pipes a remote script
Besides the shell install line the checks already flagged, the README's Windows command is `powershell -c "irm https://pi.dev/install.ps1 | iex"` — the same pattern of running a script fetched from pi.dev, which may also install Node.js. The alternative given is `npm install -g --ignore-scripts @earendil-works/pi-coding-agent`, which the README says does not pin transitive dependencies.
AI review
- Infodownloads model data from pi dev while running
README, Run with Nix: "Nix builds are offline, so the bundled model data comes from a pi.dev model catalog revision pinned in nix/model-catalog.json. At runtime, Pi still overlays newer catalog data from pi.dev as usual." The program therefore contacts the project's own server during use, whichever way you installed it.
AI review
- Infonew issues and pull requests are auto closed
The first line of the README: "New issues and PRs from new contributors are auto-closed by default. Maintainers review auto-closed issues daily." With 119189 stars and 314 open issues, a problem you report may be closed before a person reads it.
AI review
- Infoproject asks users to publish their sessions
README, "Share your OSS coding agent sessions": the author asks users to publish their agent sessions as public data on Hugging Face with the separate `badlogic/pi-share-hf` tool, so "public OSS session data helps improve coding agents". It is a separate manual step, but those sessions contain your prompts, code and errors.
AI review
Evidence is shown in its original language.
What Radar could not check
- How much the AI provider charges for real work, which depends on the model and task size
- What the agent actually does to your files and what commands it runs, since nothing asks for permission
- Whether the project's telemetry package sends anything by default, which only running it would show
- Whether the installer's Node.js install clashes with a Node version you already have
- Disk use after installing, estimated at 43 to 129 MB before model data
- Whether extensions and packages you install from npm or git are what they claim to be