Developers can run autonomous AI agents in a persistent desktop workspace
This tool moves AI coding agents out of the terminal or IDE into an independent desktop environment where projects, sessions, and plugins coexist. It manages the process of agents reading code, modifying files, executing commands, and delegating tasks to other agents within a single persistent workspace.
Workspace structure extended by plugins
The core is a plugin system that extends both the agents and the desktop environment. Beyond adding tools or skills for agents, you can attach panels, floating widgets, themes, background services, and MCP servers to the desktop itself. For example, a voice agent plugin bundles a floating widget, voice service, agent tools, and commands into a single package. Plugins are distributed as .piplug packages or installed via a marketplace, allowing the workspace to be transformed into various forms such as development tools, voice agents, or DevOps consoles.
Three work modes and multi-agent collaboration
There are three ways to assign work to agents. The 'Agent' mode has the agent directly read, modify, and execute code for routine development tasks. The 'Plan' mode requires reviewing an implementation plan before execution for refactoring or high-risk changes. The 'Goal' mode fixes objectives and acceptance criteria for complex long-term tasks, allowing the agent to choose and execute the path. Since complex tasks may not fit in a single context window, you can delegate work to sub-agents with independent contexts or use a session orchestrator to coordinate parallel worker sessions, which are full sessions.
Model independence and local-first data management
The system is not dependent on a specific model vendor, allowing free switching between various cloud and local models such as OpenAI, Anthropic, Ollama, and LM Studio. You can use different models for different sessions or switch models between planning, coding, and review stages within the same session. Data is stored locally by default, with projects, sessions, settings, and logs all remaining on the local machine. API credentials are stored in the OS keychain, and no telemetry is collected. Context is sent directly to the provider only when using remote models.
Permission control and session persistence
When agents perform file reading, code editing, command execution, tool calls, or task delegation, privileged operations must pass through a permission hierarchy. Users can set the autonomy for each session to allow, confirm, or deny. Sessions are organized by project, session, agent, and task, ensuring they persist after an app restart rather than being one-off chats. You can pin, archive, branch, and search sessions, and exercise fine-grained control such as queuing prompts while an agent is running or referencing project files with @.
Pre-use checks during the initial preview stage
Linux users require glibc 2.35 or higher, with support for Ubuntu 22.04+, Debian 12+, and Fedora 36+. macOS releases have undergone Apple Developer ID signing and notarization. Building from source requires Node.js 22.19 or higher, pnpm 10 or higher, and a stable Rust toolchain. You can import local sessions from existing Claude Code, Codex, OpenCode, and Pi, which is useful for continuing previous work. The license is GNU Lesser General Public License v3.0.
