normal-computing-oss/fuji-web

Fuji is an AI agent that lives in your browser's sidepanel. You can now get tasks done online with a single command!

Fuji‑Web — AI‑powered browser‑automation extension

What it is : Fuji‑Web is a Chrome (and compatible) extension that lets a large‑language‑model (LLM) act as a “co‑pilot” inside the browser. You type a natural‑language task (e.g. “book a flight from NYC to Paris for next Friday”), the extension sends the prompt – together with any page screenshots – to the LLM via your own OpenAI or Anthropic API key, receives a plan, and then drives the page automatically while showing you each step.

How it works

  • The UI lives in a side‑panel opened by a Fuji icon. You paste an API key; the key is stored locally in the browser and never leaves your machine.
  • When you submit a task, the extension captures the current page (text + image) and forwards it to the selected LLM.
  • The model returns a series of actions (click, type, select, scroll, etc.). Fuji‑Web executes those actions through the Chrome extension APIs, updating the side‑panel with a human‑readable explanation of every step.
  • All communication is a direct call to the OpenAI/Anthropic endpoint – there is no backend server.

Getting started

  1. Install – download the latest fuji‑extension.zip from the releases page, unzip it, and load it as an unpacked extension via chrome://extensions (enable Developer mode).
  2. Configure – open the side‑panel, paste your OpenAI or Anthropic API key.
  3. Run – navigate to any site, type a task, and watch the assistant act.

Building from source

  • Requires Node 20 (or a recent lower version) and pnpm.
  • Clone the repo → pnpm install → pnpm dev for hot‑reloading or pnpm build to produce the dist folder that you load into Chrome.

Roadmap highlights

  • Public API for integration with automation tools like Puppeteer, Playwright, Selenium.
  • More complex, multi‑tab workflows and richer UI interactions (dropdowns, full‑page extraction).
  • Saving and sharing of workflows, plus a community‑driven knowledge base to improve the model’s performance.

Why it matters Fuji‑Web demonstrates a practical, user‑controlled way to harness LLMs for everyday web tasks without a cloud‑hosted service. By keeping the API key in‑browser and exposing the action plan, it gives users transparency and privacy while showcasing how LLM‑driven agents can be embedded directly into the tools we already use.

Further reading

All details above are taken directly from the repository’s README.

Related

  • Project
    neo4j-labs/agent-memory

    A graph-native memory system for AI agents that uses Neo4j to store conversations, build knowledge graphs, and record reasoning traces.

  • Project
    sleekdotdesign/agent-skills

    A collection of AI agent skills that enable agents to design mobile apps, create screens, and manage projects using the Sleek design tool.

  • Project
    selfimproving-agent/Awesome-Self-Improving-Agents

    This is a curated, continuously evolving hub that collects papers, benchmarks, blogs, podcasts, and other resources on self-improving agentic systems. It organizes the field into two main branches—improving the underlying model and improving the surrounding scaffolding—and serves as a practical guide for researchers and practitioners.

  • Project
    AMA-CMFAI/LAMBDA

    A data analysis agent that converts natural-language questions into reproducible analysis workflows, including code execution, visualization, and report generation.