RunanywhereAI/on-device-browser-agent

On-device AI browser automation using WebLLM. No cloud, no API keys, fully private.

What it solves

It eliminates the need for cloud API keys and external servers when using AI web agents. By running the model directly on the user's machine, it ensures that page content and prompts never leave the browser, providing a private, free, and zero-config way to automate web tasks.

How it works

This project is a fork of Nanobrowser that replaces cloud-based LLM calls with on-device inference. It uses the RunAnywhere SDK, leveraging WebAssembly and WebGPU to execute models inside the Chrome extension itself. While it defaults to local execution, it retains the ability to connect to cloud providers for tasks that require more powerful frontier models.

Who it’s for

Users who want a private AI browser agent that runs locally without subscription costs or API keys, and developers interested in on-device AI execution via WebGPU.

Highlights

  • Local-first inference: Runs models inside the extension using WebAssembly and WebGPU.
  • Privacy-centric: Page data and prompts stay on the local machine by default.
  • Zero configuration: Automatically selects a model based on the user's hardware.
  • Hybrid flexibility: Supports both local models and optional cloud-based LLM providers.

Related

  • Project
    Priler/samurai

    An AI-powered auto-moderator bot for Telegram that uses ML models to detect spam, profanity, and NSFW content while managing user reputation.

  • Project
    FradSer/mcp-server-mas-sequential-thinking

    An MCP server that enhances LLM clients with a multi-agent system for sequential thinking, processing each thought through a team of specialized AI agents for multi-dimensional analysis.

  • Project
    ertugrulakben/cashclaw

    CashClaw is an npm‑distributed pack of 13 OpenClaw “skills” that let an AI agent run a freelance‑style business: it can find jobs on the HYRVE AI marketplace, quote, execute services (SEO, content, lead‑gen, etc.), invoice via Stripe and collect payment. The bundle also includes a Guard layer that enforces hard cost caps, recursion limits and tool deny‑lists, with optional Telegram alerts. Installation is a one‑liner (`npx cashclaw init`), after which the agent can poll HYRVE, auto‑accept gigs, and operate autonomously while staying protected from runaway LLM spend.

  • Project
    heaven999b/hello-agent-system

    An open-source, bilingual (Chinese/English) training bootcamp with 32 lessons that takes you from calling LLM APIs to designing and operating production-grade agent systems. You hand-write every layer—tools, concurrency, reliability, security, observability, evaluation, distributed execution—then swap in mature components like Postgres, Redis, and Temporal, all with automated tests and real process fault injection.