LIVE
+++ Commentary: AI firms grab training data at any cost +++ EdgeEver puts an MCP server inside your note archive +++ PiClaw puts the Pi coding agent in a browser workspace +++ LLMs as a Cognitive Virus: Preprint Models Dependence +++ AIHawk: an open-source browser agent that speaks MCP +++ OpenGUI: an Android agent that taps through real apps ++++++ Commentary: AI firms grab training data at any cost +++ EdgeEver puts an MCP server inside your note archive +++ PiClaw puts the Pi coding agent in a browser workspace +++ LLMs as a Cognitive Virus: Preprint Models Dependence +++ AIHawk: an open-source browser agent that speaks MCP +++ OpenGUI: an Android agent that taps through real apps +++
All news ›
AI IN LIFE AI IN LIFENEWS
DAILY
DE EN
TOOLS

PiClaw puts the Pi coding agent in a browser workspace

Chat, file browser, editor and terminal share one window; the MIT-licensed project is explicitly built for a single user, not a team.

PiClaw puts the Pi coding agent in a browser workspace

Symbolic image: a compact home server with blinking status LEDs beside a monitor where abstract color blocks shift across three panes.

PiClaw is an MIT-licensed, self-hosted AI workspace that puts the Pi coding agent, a file browser, an editor, and a terminal into one browser window built for exactly one user.

At a glance

  • MIT license, TypeScript codebase; install via Docker, portable release, Bun repo checkout, or source build with experimental desktop shell.
  • The Docker quick start binds the web server to 127.0.0.1:8080 and mounts one directory each for config and workspace.
  • Built on the Pi Coding Agent from badlogic/pi-mono; the project states no direct affiliation with pi.dev and no Raspberry Pi link.
  • Single-user only: extra tabs or conversations are not separate accounts, and multi-user modes remain gated development work.
  • As of September 7, 2026: about 850 stars, 80 forks; interface available in English, Simplified Chinese, and Japanese.

PiClaw is a self-hosted AI workspace for one person, where a coding agent, a file browser, an editor, and a terminal all live in the same browser window. The repository rcarmo/piclaw ships under the MIT license.

One window instead of a toolchain

The project page frames PiClaw as a place to work with an agent, edit files, run commands, and inspect what came back without leaving the page. Chat streams as it generates, the model can be picked per conversation, and a run can be steered while it is still going. Conversations persist and stay searchable, and a file-based memory system carries context between sessions.

Results get dedicated viewers rather than a wall of raw text: CSV and TSV tables, PDF, images, video, code, plus VNC panes for remote displays. Scheduled tasks, skills, MCP server integration, browser automation, and image processing fill out the rest of the feature list. The interface is offered in English, Simplified Chinese, and Japanese.

The Pi here is not a Raspberry Pi

The name points at the Pi Coding Agent from the badlogic/pi-mono repository, which PiClaw is built on top of. The project says plainly that it is not directly affiliated with pi.dev and describes itself as a derivative work on the Pi core. Nothing on the page connects the tool to single-board computers.

Getting it running

Docker is the recommended route. Portable releases cover Linux and Apple Silicon macOS, with Windows marked experimental, and there is also a Bun install straight from the repository plus a source build with an experimental desktop shell. The quick-start command binds the web server to 127.0.0.1:8080 and mounts one directory for configuration and one for the workspace.

What you supply is Docker plus credentials for a model provider, or a local model server the container can reach. Local OpenAI-compatible servers, llama.cpp, and Azure OpenAI come up by name, and the active model is configured from inside the chat.

Built for one, and it means it

The sharpest constraint is a design decision rather than a missing feature: PiClaw targets single-user deployments. Opening a second tab or starting another conversation does not create a separate account, and multi-user handling is described as gated development work. The quick-start command binding the port to the loopback address follows from that, since exposing the interface publicly would hand strangers an agent with terminal access to the workspace.

What this write-up cannot confirm

Everything above comes from the project's own GitHub page, and no independent second source was available at this stage. That leaves performance, resource requirements, and the real-world stability of the experimental Windows and desktop builds unverified. The reach figures are a snapshot too: roughly 850 stars and 80 forks on September 7, 2026.

◈ AI-GENERATED REPORT · SOURCES LINKED

FAQ

What is PiClaw?

A self-hosted, MIT-licensed web interface for the Pi coding agent, combining streaming chat, a file browser, an editor, a terminal, and viewers for CSV, PDF, images, video, and code in a single browser window.

Can several people share one PiClaw instance?

No. The project supports single-user deployments only, and it warns that separate conversations or browser tabs are not separate accounts. Multi-user handling is still gated development work.

Does PiClaw work with a local model?

Yes, according to the project page: it needs credentials for a model provider or a reachable local model server, and it names local OpenAI-compatible servers, llama.cpp, and Azure OpenAI. The quick start serves the interface on 127.0.0.1:8080.