PI-Desktop is a free, open-source desktop workspace for AI coding agents that lets a team bring its own model instead of paying $20 to $100 a seat for Cursor or GitHub Copilot.
A five-person dev shop paying for Cursor seats for every engineer is looking at $20 to $40 a month per person, on top of whatever model usage runs over the included limit. A team that wants the same agentic coding workflow without the recurring seat cost now has a free option to point at: PI-Desktop, an open-source desktop workspace for AI coding agents that shipped its latest release the same day this was written.
PI-Desktop is not a hosted product. It is a downloadable application for macOS, Windows, and Linux that gives a coding agent its own workspace instead of tucking it into a terminal or an editor extension. You open a local project, connect a model provider of your choice, and the agent reads files, edits code, and runs commands from there. The company behind it does not put itself in the middle of that conversation. There is no PI-Desktop account requirement and no relay service the requests have to pass through. Model requests go straight from your machine to whatever provider you configured, whether that is OpenAI, Anthropic, a local model through Ollama or LM Studio, or any OpenAI-compatible API.
That bring-your-own-model design is the whole cost story. Cursor's Individual plan runs $20 a month per user, and its Teams plan is $40 a month per user, according to Cursor's own pricing page. GitHub Copilot's Pro plan is $10 a month per user, with Pro+ at $39 and Max at $100 for heavier agent usage, per GitHub's plans page. Those prices bundle the model access in. PI-Desktop charges nothing for the application itself; a team instead pays only for whatever model usage it actually consumes, at whatever provider's own rates, and can mix models or swap providers per session without recreating anything.
The project is licensed LGPL-3.0, which permits commercial use, and it is not a weekend experiment. The repository shows 55 tagged releases, with the newest, v0.14.5, published the day this article was written. It has picked up 1.6 thousand stars and 151 forks. The README lists Agent, Plan, and Goal modes (do the work directly, wait for an approved plan first, or work toward an approved outcome), support for subagents that handle research or multi-file changes in the background, and extension points for MCP servers, Skills, and installable plugins. Teams already using Claude Code, Codex, OpenCode, or Pi can import their existing local sessions rather than starting over.
Who actually picks this up. An internal engineering lead or a technical co-founder evaluating tool spend across a small team is the natural first user, since the savings scale with headcount and the setup work (installing an app, connecting a provider account) is a one-time cost rather than a recurring one. A solo founder who already pays for an API key somewhere and doesn't want a second subscription layered on top fits too. It is a harder sell for a non-technical operator who wants support tickets and a vendor to call when something breaks; that role does not really exist here.
The honest caveats matter as much as the pitch. PI-Desktop labels itself Early Preview, and the maintainers say APIs and desktop behavior may still change. The tagged macOS build is unsigned by default, which means a manual step to open it the first time, unless a team specifically requests a signed and notarized build. Linux support requires a fairly recent glibc, so Ubuntu 20.04 or Debian 11 and older cannot run the bundled host. And self-hosting the workspace does not make the model free; a team still pays its provider directly for every token, it just is not bundled into a seat price and is not marked up by a middleman. Whether that is cheaper depends entirely on how much a team's engineers actually use the agent, and PI-Desktop's own docs do not publish a usage benchmark to compare against.
The interesting part of PI-Desktop is not that it is free. Plenty of things are free and worse than the thing they are replacing. It is that the cost structure moves from a fixed per-seat number to a variable one tied to actual usage, and that only makes sense once a team has enough visibility into its own model spend to know which side of that trade it is really on.