Yes: the README provides a complete command sequence from cloning the repository to starting the development environment, and its runtime requirements match Node.js 24, pnpm 11, and the stable Rust toolchain.

• The Requirements section requires Node.js >=22.19, pnpm >=10, and a stable Rust toolchain. It also states that the repository currently pins pnpm 11 while CI and release builds use Node 24.
• The Run locally section provides the sequence git clone, pnpm install, cargo build -p host-core, pnpm build:js, and pnpm dev for local setup and startup.
• The architecture contains a React Renderer, Electron Main, Rust Host Core, and pi Agent Sidecar. Local construction therefore includes compiling the Rust host core, not only bundling TypeScript.
• Although TypeScript is the main language, the project data also reports substantial Rust source volume, confirming that Rust is a real build dependency rather than an optional component.

The README does not provide platform-specific native dependencies, troubleshooting for build failures, or a complete CI-reproduction procedure.

Yes: Plan mode directly separates repository research from execution, while privileged tools remain behind the permission layer.

• The README’s Agent, Plan, and Goal table states that Plan studies the repository, writes a frozen implementation plan, and waits until the user signs off before execution starts.
• It specifically positions Plan for changes that are large or risky when the user wants to approve the approach first, matching the stated refactoring constraint.
• After execution, the workspace provides a Review panel for inspecting edits, command output, and application previews, rather than limiting review to the model’s textual response.
• The README also says privileged tools go through the permission layer in every mode, so Plan is not an unrestricted execution path.

The limitation is governance scope: the available material does not promise organization-wide audit trails, CI gates, or centralized approval records. Those controls must remain in the existing engineering process.

Yes: the README explicitly provides import support for local sessions from Claude Code, Codex, OpenCode, and Pi, although full-fidelity compatibility still needs verification.

• The “Import your existing sessions” section lists all four supported tools, so PI-Desktop is not limited to newly created conversations.
• The documented entry point is Settings → Import, allowing existing work to be brought into the desktop workspace.
• New conversations are stored locally as JSONL and indexed with SQLite, matching the product’s focus on long-lived, navigable session management.
• The README also allows providers, models, and reasoning levels to be switched directly from the Composer, so imported work can continue with a different model configuration.

Import does not necessarily mean that every external state is reproduced. The README does not define migration coverage for attachments, tool-call results, extension settings, branch state, or provider-specific metadata, so it should not be treated as a complete environment clone.

It depends: PI-Desktop can connect to a local endpoint, but “local-first” does not automatically mean fully offline.

• The README explicitly supports local gateways such as Ollama and LM Studio, as well as OpenAI-compatible APIs. Model requests are sent directly to the provider or endpoint configured by the user.
• Conversations are stored locally as JSONL with a SQLite index; settings, logs, and credentials remain local or in the operating-system keychain. There is no required account, mandatory PI relay, or PI-Desktop telemetry.
• If a remote provider is selected, the required project context and tool results are still sent to that provider. The application also does not remove local-model limits around context length, reasoning quality, GPU capacity, or tool-call compatibility.

It is suitable when you can enforce a local endpoint, but it is not itself a proof of complete offline operation.

Yes: PI-Desktop places projects, sessions, file changes, reviews, command output, and previews in one desktop workspace, and it ships a macOS Apple Silicon build.

• The Desktop-first section says users can work across repositories and sessions without tying the agent workflow to one editor or terminal.
• The From prompt to patch flow supports opening any local repository or project directory, then reviewing edits, command output, and application previews in the same interface.
• The Download table lists .dmg and .zip packages for macOS Apple Silicon, directly matching the hardware constraint.
• The project data identifies TypeScript as the main language, while the architecture uses Electron, React, and a Rust Host Core, confirming that this is an independent desktop application rather than an editor plug-in.

The qualification is maturity: the current release is v0.14.6-rc.4, and the project is an Early Preview. Platform behavior, session recovery, and upgrades may be less predictable than in a mature editor, and projects dependent on internal IDE state may still require the original editor.

It depends: PI-Desktop offers layered extension points and a permission layer, but the README does not prove that every third-party extension has strong sandboxing or uniform security review.

• The Extensions without rebuilding the app section says Skills provide reusable instructions and workflows, and can be installed globally or activated per project.
• The project insights identify MCP, Subagents, and the plugin marketplace as additional extension layers, allowing tools, prompts, and task delegation to be composed.
• The Architecture section assigns permissions, filesystem access, SQLite, and secrets to the Rust Host Core, while the renderer has no Node integration. This establishes a clearer boundary around privileged operations.
• The Agent, Plan, and Goal table says privileged tools go through the permission layer in every mode.

The risk is the extension itself: the project insights warn that untrusted or over-privileged plugins, MCP servers, and Skills may access code, execute commands, touch credentials, or exfiltrate data. It is suitable for controlled extensions, not for blindly trusting arbitrary plugins.