w01fb0ss/PiBot
AI EngineeringA local-first desktop app where multiple named AI agents collaborate in groups, use tools and await approval for risky actions.
This project ships 37 agent skillsSkill list and install →Overview
PiBot is a desktop application for running a team of named AI agents on your own machine. You can chat with one agent, or place several in a group where a coordinator plans work, distributes steps and collects results. Agents can plan, run commands, edit files in a chosen working folder, browse the web and operate the computer; every risky step goes through an approval policy. The agent kernel is pi, pinned as an npm dependency, and the non-Electron "brain" packages can run headless.
Agents are configured with a name, role (developer, researcher, writer, translator, etc.), model, skills, permissions and memory. Built-in role and group definitions support offline use, with a larger role library in pibot-roles. Other features include scheduled or folder-triggered routines, MCP servers and app connections, and optional remote use through an Android app, headless runners, a self-hostable relay, and editor access over ACP.
Getting started
Install the desktop app from the official download page and follow the getting-started guide.
To build from source, install Node matching .nvmrc and pnpm 12, then:
git clone https://github.com/w01fb0ss/PiBot.git
cd PiBot
nvm use
pnpm install
pnpm dev
pnpm dev starts the desktop app in development mode. Set PIBOT_DATA_DIR to keep development data separate from an installed copy. pnpm check runs install, typecheck, lint, policy-matrix and test checks, while pnpm verify additionally builds and runs renderer smoke and end-to-end acceptance suites. The source build has no installer target; use pnpm build for an electron-vite build.
Requirements and constraints
The desktop app supports macOS 13 Ventura or newer, Windows 10/11 and 64-bit Linux; the mobile client is Android. Commands run inside @anthropic-ai/sandbox-runtime, and agent actions, approvals and outbound requests are recorded in an audit log. The project states that it has no telemetry and its own servers do not receive chats or files; model prompts go to the provider you configure. Separate outbound connections occur for update checks, websites and searches used by agents, and optional features such as phone, runners and connections, documented in the outbound-connections notes.
Models can be local via Ollama or remote through OpenRouter, your own API keys, or supported subscription sign-ins. The project is Apache-2.0 licensed. Repository layout splits the Electron desktop app, Android app, relay, headless runner, CLI, bundled plugins and the multi-layer agent packages; see AGENTS.md for the brain's architecture and CONTRIBUTING.md before contributing.
HotRepo's take
The repository entered HotRepo's new-project list on 2026-11-21, its first recorded day, with 116 stars and 16 forks gained in a roughly 10-hour window. That is enough to say the launch attracted early attention, but one snapshot cannot establish a trend. The codebase supports the product story rather than just pointing at a website: separate desktop, mobile, relay, runner and CLI apps, 125 skill files, and package-level tests make it a real multi-agent workspace. The most concrete differentiators to inspect are the shared approval policy engine, sandboxed command execution and audit logging, since those determine how much autonomy the agents actually have.
Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.
You might also like
- totec448-spec/chat-on-steroidsA local MCP workspace that lets ChatGPT read files, run commands and delegate tasks across multiple agents.
- agentes-chat/agentes-chatSelf-hosted team chat where people and AI agents share channels, DMs and approvals.
- Verlintas/NovaBAICLocal-first Android AI agent that calls 58 device tools to read the screen, automate UI, manage files and run scheduled tasks.
- androoAGI/starnetA local-first desktop harness where multiple AI agents run concurrently in separate workspaces, with the pixel-art station reflecting real runtime state.
- tinyhumansai/openhumanA Rust agent harness aimed at running hundreds of LLM agents cheaply on one small server, usable as a library or desktop app.