PI‑Desktop – A modular, local‑first desktop for AI agents

What it is

PI‑Desktop is a cross‑platform (macOS, Windows, Linux) desktop application that gives AI agents a persistent, extensible workspace. It is not an IDE plug‑in or a thin wrapper around a single LLM; instead it provides a desktop platform where you can run, orchestrate, and manage agents, plug‑ins, and model back‑ends.

Core ideas

Idea	What it means
Independent workspace	Projects, sessions, and agents live in their own sandbox, separate from any IDE or terminal.
Plugin‑powered	Functionality is added through plug‑ins that can contribute panels, commands, widgets, services, themes, or new agent tools.
Agent orchestration	One agent can spawn sub‑agents (lightweight background workers) or full Worker Sessions that run in parallel, each with its own context.
Model freedom	You can point the runtime at OpenAI, Anthropic, Ollama, LM Studio, custom gateways, or any local model. Models are interchangeable per session or even mid‑session.

How you work with it

1. Download the app (DMG, installer, AppImage, etc.).
2. Connect a model by adding a provider (API key, endpoint, etc.).
3. Open a local project (any git repo or folder).
4. Choose a workflow mode:
   • Agent – give a concrete task and let the agent code, run tests, edit files.
   • Plan – the agent first produces an implementation plan before touching code.
   • Goal – you set an outcome and acceptance criteria; the agent decides the path.

All work is stored locally: projects, sessions, logs, and settings never leave your machine unless you explicitly send a request to a remote model.

Plugins – the extensibility surface

Plugins are first‑class citizens and can add any of the following capabilities:

• Commands (global actions)
• Panels (stand‑alone UI windows)
• Floating widgets (e.g., voice orbs, status lights)
• Work‑panel views (right‑side workspace tabs)
• Agent tools / skills / completions (functions the agent can call)
• Themes
• MCP servers (local/remote services the desktop can talk to)
• Background services
• Message bus for inter‑plugin communication

Plugins are distributed as .piplug packages or via the built‑in marketplace.

Session model

• Project → Session → Agent → Work hierarchy.
• Sessions persist across app restarts, can be archived, branched, searched, and queued.
• You can reference files with @, use slash commands, view diffs, and inspect command output.
• Subagents and Worker Sessions give you parallelism without overloading a single LLM context window.

Model‑agnostic architecture

You configure each model with:

• Provider, model ID, context window, temperature, etc.
• Different sessions can use different models, and a single session can switch models on the fly (e.g., planning with Model A, coding with Model B).

Local‑first & privacy

• All data (projects, sessions, logs) stays on your device.
• API credentials are stored in the OS keychain.
• No mandatory account, telemetry, or relay service.
• Permission layer lets you approve or deny privileged actions (file writes, command execution, etc.).

Development & extensibility

• Built on the pi ecosystem (pi-ai, pi-agent-core, pi-coding-agent).
• Source code requires Node ≥22, pnpm, and a stable Rust toolchain.
• Quick start: git clone … && pnpm install && cargo build -p host-core && pnpm dev.
• Documentation includes architecture specs, plugin development guides, and an E2E test plan.

Who might use it?

• Developers who want a dedicated AI‑assistant desktop rather than a chat window.
• Teams building custom AI‑agent products (voice agents, DevOps consoles, data assistants) via plugins.
• Researchers experimenting with multi‑agent orchestration and model swapping.

License

GNU LGPL v3 – you can use, modify, and redistribute the core while keeping it open‑source.

---

Bottom line: PI‑Desktop is a genuine, open‑source platform that lets you run, manage, and extend AI agents on your own machine, with a strong focus on modularity, model‑agnosticism, and privacy.