PI-Desktop Gives Your AI Agents a Place to Actually Live
You've probably got a terminal agent running tasks in one window, an IDE extension handling completions in another, and a browser tab open for the model's web UI. It works, but nothing talks to anything else, and your workflow is scattered across tools that don't know each other exist. PI-Desktop is an attempt to fix that by giving agents their own desktop workspace—one that isn't bolted onto an editor or a shell.
What It Does
PI-Desktop is a modular desktop application for AI agent workflows, available on macOS, Windows, and Linux. The core idea is simple: instead of agents being a feature inside some other tool, they get a persistent workspace of their own. Projects, sessions, reviews, previews, and agents all live together in one environment.
The architecture is deliberately plugin-driven. The core stays small, and your actual workflow gets assembled through extensions. Plugins can add things at three levels: agent capabilities (tools, skills, completion, "pi Extensions"), workspace features (commands, panels, work panel views, floating widgets, themes), and platform services (MCP servers, resident services, a plugin message bus). It's model-agnostic, so you can use cloud models, local models, custom gateways, or compatible APIs without rebuilding your setup when you switch.
The current release line is 0.15.x, labeled as an early preview.
Why It's Cool
Plugins aren't an afterthought here—they're the whole point. A lot of tools treat extensions as a nice-to-have. PI-Desktop seems to treat them as the primary way you build your environment. The README makes a point of saying a plugin doesn't have to be "just another tool"—it can be an entire product. The examples given are telling: a Voice Agent plugin could bundle a floating widget, a speech service, an agent tool, and commands. A GitHub Workspace plugin could be a work panel, an MCP server, agent tools, and a background service. That's a fundamentally different mental model than "here's a function you can call."
Agent orchestration is built in. The README mentions delegating to subagents or coordinating full worker sessions in parallel. One agent isn't always enough, and having that coordination live in the same workspace rather than scattered across separate processes is a real practical advantage.
The workspace is independent by design. This isn't an IDE extension and it isn't a terminal wrapper. Projects, sessions, and agents persist in their own environment. If you've ever lost context because you closed a tab or your editor crashed, you'll understand why that matters.
Model freedom is treated as a first-class concern. The tagline is "Your models stay replaceable." Switching between a local model for sensitive work and a cloud model for heavier tasks shouldn't require rearchitecting your workflow, and PI-Desktop seems to take that seriously.
The plugin capability table is genuinely broad. Commands add actions to a global command system. Panels open standalone plugin interfaces. Floating widgets let you build voice orbs, status lights, timers—things that live on your desktop rather than inside a chat window. That's a level of UI extensibility you don't see often.
How to Try It
The fastest path is to grab a prebuilt release:
- Head to the releases page and download the build for your platform (macOS, Windows, or Linux).
- Install and launch it. The README describes it as local-first, so your projects stay on your machine.
- If you want to extend it, the plugin development docs are linked directly from the README.
- For more detail on configuration and usage, the documentation site is the place to go.
If you'd rather build from source, the repository is at github.com/vastsa/PI-Desktop. There are also screenshots if you want to see what it looks like before committing to a download.
Final Thoughts
PI-Desktop is in early preview (0.15.x), so it's not the kind of thing you'd bet a production workflow on just yet. But the design philosophy is coherent and worth paying attention to. If you're someone who's been stitching together agents, models, and tools across half a dozen applications, this is aimed squarely at you. The plugin architecture means the ceiling here depends on what the community builds, and the foundation looks solid enough to find out. Worth a download if you're curious about where agent tooling might be heading.