PI-Desktop
A modular desktop workspace for AI agents
Bring projects, agents, models, plugins, and workflows into one persistent desktop environment.
Local-first · Model-agnostic · Plugin-powered · macOS / Windows / Linux
Download · Documentation · Build a Plugin · Screenshots · 简体中文
Your projects stay local · Your models stay replaceable · Your workspace stays yours
Current release line: 0.17.x (Early Preview).
Why PI-Desktop?
Terminal agents are great at execution. IDE agents are great at living inside an editor.
PI-Desktop goes one step further:
Give AI agents a persistent, independent, and extensible desktop workspace of their own.
Independent Workspace
No dependency on a specific IDE or terminal.
Projects, sessions, reviews, previews, and agents all live in their own workspace.
Plugin-Powered
Plugins extend more than the agent.
Add panels, views, widgets, tools, MCP servers, themes, and background services.
Agent Orchestration
One agent is not always enough.
Delegate to Subagents or coordinate full Worker Sessions in parallel.
Model Freedom
Cloud models, local models, custom gateways, compatible APIs.
Switch models without rebuilding your workflow.
It is not a wrapper around one model. It is not another IDE extension.
It is a desktop platform for agent workflows.
Plugins are part of the workspace, not an afterthought
PI-Desktop keeps the Core focused.
Your actual workflow is assembled through extensions.
Agent
Extend what the agent can do
Agent Tools Skills Completion pi Extensions
Workspace
Extend the desktop itself
Commands Panels Work Panel Views Floating Widgets Themes
Platform
Extend the runtime
MCP Servers Resident Services Plugin Message Bus
A plugin does not have to be “just another tool.”
It can be an entire product:
Voice Agent
├── Floating Widget
├── Speech Service
├── Agent Tool
└── Commands
GitHub Workspace
├── Work Panel
├── MCP Server
├── Agent Tools
└── Background Service
Session Analytics
├── Dashboard
├── Commands
└── Workspace View
What can a plugin add?
| Capability | What it enables |
|---|---|
| Command | Add actions to the global command system |
| Panel | Open a standalone plugin interface |
| Floating Widget | Build voice orbs, status lights, timers, and other floating UI |
| Work Panel View | Add new views to the right-side workspace |
| Agent Tool | Register tools callable by the agent |
| Completion | Use the models already configured by the user |
| Skill | Add reusable agent capabilities and workflows |
| Theme | Customize workspace appearance |
| MCP Server | Connect local or remote MCP servers |
| Service | Run persistent background work |
| Message Bus | Let plugins communicate with each other |
Plugins can be distributed as .piplug packages or installed through the marketplace.
Build your first plugin →
One foundation, many workflows
PI-Desktop
│
┌──────────────────┼──────────────────┐
│ │ │
Agent Workspace Platform
│ │ │
Agent Tools Panels MCP
Skills Widgets Services
Subagents Views Message Bus
pi Extensions Themes
│ │ │
└──────────────────┼──────────────────┘
│
Your Workflow
PI-Desktop can simply be your coding agent.
Or you can turn it into:
AI Development Workspace · Voice Agent · DevOps Console · GitHub Workspace · Data Assistant · Multi-Agent Control Center · Automation Platform
The Core provides the foundation. Plugins decide what your workspace becomes.
Three ways to work
Agent
Give it a task. Let it work.
Read code, edit files, run commands, test, and iterate.
Best for day-to-day development.
Plan
Review the approach before execution.
The agent studies the project first and produces an implementation plan.
Best for refactors and high-risk changes.
Goal
Define the outcome. Let the agent choose the path.
Lock the objective and acceptance criteria, then let the agent drive execution.
Best for complex and long-running tasks.
Privileged operations still pass through PI-Desktop's permission layer.
When one agent is not enough
Complex work should not be forced into one context window.
PI-Desktop provides two levels of delegation.
Subagents
Delegate independent work to background agents:
Code exploration · Implementation · Test analysis · Research · Review
Each Subagent gets its own context and reports the result back to the parent agent.
Session Orchestrator
For longer-lived work, delegate to full Worker Sessions.
Main Session
│
├── Worker A
│ └── Frontend
│
├── Worker B
│ └── Backend
│
├── Worker C
│ └── Tests
│
└── Worker D
└── Review
Workers are full PI-Desktop sessions:
Independent context · Independent execution · Directly inspectable · Reusable · Full transcript
Move from “one agent helps me code” to “multiple agents divide and complete the work.”
Built for work that lasts
PI-Desktop is organized around:
Project → Session → Agent → Work
—not around disposable chat threads.
You can:
- Manage multiple projects and sessions
- Pin, archive, branch, and search sessions
- Queue prompts while an agent is already running
- Reference project files with
@ - Use slash commands
- Review diffs
- Inspect command output
- Work with the right-side Work Panel
- Keep streaming checkpoints
- Recover interrupted work whenever possible
A Session can continue across multiple app launches.
See what the agent is doing
Swap the model, keep the workflow
PI-Desktop does not tie your workflow to a single model vendor.
Use:
OpenAI · Anthropic · OpenAI-Compatible APIs · Custom Gateways · Ollama · LM Studio · Local Models
Configure each model independently:
Provider · Model ID · Context Window · Output Limit · Reasoning / Thinking · Temperature · OAuth · API Key · Endpoint
Different Sessions can use different models.
The same Ses