PiBot ships a local-first team of desktop AI agents
A new Apache-2.0 TypeScript project puts named AI bots in a group chat on your own machine, with an approval gate in front of every risky command.
PiBot is an open-source, local-first AI agent team for the desktop, published as w01fb0ss/PiBot on GitHub. As of 22 November 2026 the repository had 301 stars and 44 forks, collected in the one day since it was created on 21 November 2026. It is written in TypeScript and released under Apache-2.0.
Named bots that share a working folder
PiBot is a desktop app, not a hosted service. You create Bots with a name and a role such as developer, researcher, writer or translator, and give each one its own model, skills, permissions and memory. You can chat with a single Bot or put several into a group. In a group, a coordinator plans the work, hands steps to members and collects the results, and you confirm the plan before anything runs.
Bots work inside folders you choose. The app lists every change so it can be reviewed, and on macOS changes can be restored from automatic checkpoints. Beyond chat there are routines: tasks triggered on a schedule, by a folder, or by a connection. Connections cover apps, MCP servers, skills and knowledge bases.
Approvals, sandbox and local data
Every tool call passes through one policy engine. Commands run inside a sandbox (@anthropic-ai/sandbox-runtime), and approvals, commands and anything sent outward are written to an audit log. The readme says the brain runs on your machine: chats, files, memory and keys stay local, and only the model provider you pick sees the prompts you send. Model options include local models through Ollama, OpenRouter, your own API keys and supported subscription sign-ins.
The project states it has no telemetry and that its own servers never receive chats or files. Outbound connections are limited to update checks, the sites and searches Bots use, and optional features you turn on, each explained first and listed in docs/privacy.
More than one device
Optional pieces extend the same setup: an Android app, runners on other computers or servers, a relay you can host yourself, and use from editors over ACP. Supported desktop systems are macOS 13 Ventura or later, Windows 10 or 11, and 64-bit Linux.
Building it yourself
A source build needs Node from .nvmrc and pnpm 12. pnpm dev starts the app in development, pnpm check runs a frozen install, typecheck, lint, a policy matrix and tests, pnpm verify adds a build, a renderer smoke test and end-to-end suites, and pnpm build produces an electron-vite build without an installer. Setting PIBOT_DATA_DIR keeps development data apart from an installed copy.
The repository splits into an Electron desktop app, a React Native and Expo Android app, a Cloudflare Worker relay, a headless runner and CLI, bundled plugins, and packages holding the brain. The pi agent kernel is pinned as an npm dependency and used only in packages/agent; because the brain does not depend on Electron, the same code also runs headless.
What this suggests for builders
The part worth studying is the approval boundary. An agent that edits files and runs shell commands is only as useful as the review step around it, and PiBot puts one policy engine, a sandbox and an audit log in front of every tool call. A workflow the readme supports: give a Bot a project folder, let the coordinator plan, confirm the plan, then approve the risky commands as changes appear in the list.
Limits and open questions
The star and fork counts are a one-day snapshot and will move. The readme text carries no download URL and no pricing, so it is unclear what a packaged build costs, if anything. Which subscription sign-ins are supported is not listed. Checkpoint restore is described for macOS only, and the source does not say whether Windows or Linux behave the same way. The Android app and runners are described as optional but their scope is not spelled out. Nothing in the source gives performance figures or an independent security review beyond the sandbox and policy design.
What it means for building apps with AI
PiBot is a different shape of tool from a hosted app builder: it runs on your machine and works in your folders, while a builder generates and hosts the app itself. Draftship is in the second camp, and is not affiliated with this project. You describe an app in plain language and Draftship writes the code, builds and hosts it, with a built-in database, user sign-in and hosting; paid plans add private projects, a model choice, source code access, custom domains and GitHub push. The shared lesson is the approval step. If your app calls a model, Draftship's notes on the AI chat and choosing a model cover that decision.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.