Run the pi coding agent with local or hosted models. Manage sessions, Git worktrees, code review, and web previews in one desktop application.
pi
AG-UI
CopilotKit
Angular
Electron
llama.cpp
llama-swapWe love pi for its composability and model freedom. We needed a Windows desktop client for working with local models, and wanted an experience that felt more at home on Windows than a terminal interface. That practical need became pi-kit.
pi-kit adds model setup, credential storage, session navigation, and review tools while using pi's existing session and extension formats.
We also maintain the Angular port of CopilotKit. pi-kit uses it with AG-UI to render tool interfaces, input requests, and MCP Apps in the conversation.
Read the project storyConfigure models, navigate conversation history, isolate changes in worktrees, and inspect results before merging.
Search Hugging Face and download supported GGUF models directly inside pi-kit. Review the file, verify its checksum, and add it to your llama-swap configuration—or connect a server you already run.
Read the guideNavigate pi’s real session tree, branch from an earlier turn, and keep the original approach available while you explore another.
Read the guideGive a thread its own Git branch and folder. Inspect the diff, sync with the base, and merge back when the change is ready.
Read the guideUse A2UI and sandboxed MCP Apps for rich tool interfaces, with explicit prompts when the agent needs your decision.
Read the guideUse your operating system credential store and resolver commands. Inspect the reported storage state instead of assuming every environment has a keychain.
Read the guideDraft a skill from a finished thread, replay it in a scratch worktree, and review the result before approving what gets installed.
Read the guideOpen your running project in Artifacts. You and the agent navigate, inspect, and test the same native page.
Read the guideAttach a page or element screenshot, DOM details, or selected console logs to your unsent draft.
Read the guideOpen a running web project beside the conversation. The agent can navigate, interact with, and inspect the same browser page displayed in Artifacts.
Explore the Artifacts browserStart your project server, enter its HTTP address, and navigate the same page the agent can operate. Resize the dock to check narrower layouts.
Attach a viewport, full-page, or selected-element screenshot to a message. Include DOM details or selected console entries when needed. Attachments are added to the unsent draft.
Screenshots, inspection, and logsChoose the scope that explains the issue. Use selected-element capture for a small detail or include surrounding layout with a visible-area image.
Use the page you have open, at the width where the problem happens. No need to describe a layout from memory.
Attach element details or selected console messages when the issue needs more than pixels. Review private information before sending.
Ask the agent to revisit the same flow, check the console, and capture the result. A passing build is not visual verification.
Search supported public Hugging Face GGUF models right inside pi-kit. Review the source, license, file size, and memory estimate, then explicitly download the file with checksum verification. Verified downloads are added to your selected llama-swap config.
Existing local servers can be connected without downloading another model. Hosted connections include Radius, supported subscriptions, provider API keys, and LiteLLM.
Navigate earlier turns and alternate conversation branches. Use worktrees to isolate file changes, then review unified or split diffs. Session history and Git history are separate.
Worktrees and code reviewReturn to earlier turns and inspect another approach without losing the original conversation. Navigating history does not undo filesystem changes.
Self-learning in pi-kit creates instructions for future tasks, not new model weights. The bundled @pi-kit/learn package analyzes a thread and drafts a standard pi SKILL.md file. You review the instructions before installing them.
A completed thread has the conversation, tool calls, and checks used to make a change. A recorded commit, merge, or pull-request outcome provides the base and result for verification.
Run /learn, or let the enabled attendant draft from a qualifying outcome. The analysis uses a redacted transcript and outcome diff to extract steps, encountered errors, and verification commands into a proposed skill.
Optionally run the draft in a scratch worktree at the outcome's base commit, using the first prompt. The report compares changed files and hunks and reruns recognized project checks. Reproduction can pass, be partial, or fail.
Read the proposed SKILL.md, edit the procedure, and select project or user scope. Approve installs it; Discard leaves it uninstalled. An automatically generated draft is not automatically approved.
Select the approved skill from the composer or / menu and pin it when needed. Later messages receive those instructions as context. Settings → Skills → Library keeps its origin, usage, and verification record.
The task changed the starter function and added three tests. Read opens the generated skill.
A dependency-free Node.js project with a starter function and three tests. Complete a normalization task, draft and verify a skill, then select it for a second task. The walkthrough includes the prompts, Git steps, and results to inspect.
Drafting and replay use configured models and may incur provider costs. A replay does not merge or push changes; its tools are not sandboxed to the scratch checkout. A passing comparison applies to that replay, not every future use.
API keys and connector tokens go to macOS Keychain, Windows Credential Manager, or Linux Secret Service. pi's auth.json and mcp.json only hold a command that resolves the secret when needed.
{
"openrouter": {
"type": "api_key",
"key": "!pi-kit-secret get openrouter"
}
}
Browse and install pi packages from npm, Git, or a local folder inside Settings. See the extensions, skills, prompts, and themes each package brings before you add it.
Available for macOS on Apple Silicon and Windows on x64 and ARM64.