Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

ysunyang4/webagent

AI Engineering

Browser extension that bridges web-based LLM chat pages to a local Nexus MCP server for tool calls.

111
30
JavaScript
GPL-3.0
5 days old

Project overview

WebAgent is a browser extension for Chromium-based browsers that adds a floating control panel to supported web LLM chat pages and bridges them to a locally running Nexus MCP server. Instead of API keys, it works with the web sessions you are already logged into: DeepSeek, Gemini, ChatGPT, Kimi, Qwen, Doubao, and several zhipu/Google AI Studio pages. The extension parses tool-call code blocks emitted by the model and forwards them to local MCP tools such as file operations, command execution and page snapshots, then inserts results back into the chat input.

It suits users who already run a Nexus MCP project locally and want a web LLM interface to drive it, while keeping tool permissions under user control.

Quick start

  1. Clone or download the repository.
  2. Open chrome://extensions/ (or edge://extensions/), enable Developer mode, click "Load unpacked", and select the project folder.
  3. Start Nexus locally and register a project. The default service base is http://127.0.0.1:8787 with /mcp; allowed local ports are 18080 and 21888.
  4. Open the extension popup and fill in the MCP address, the registered project ID and working directory, then test the connection. Both project ID and working directory start empty.
  5. Open a supported chat page; the floating WebAgent card appears at the top. Copy the enhanced collaboration protocol from the panel and send it to the AI yourself.
  6. Optionally enable automatic tool execution and automatic result sending in the popup. Both default to off and are independent.

After updating the extension, save drafts, reload the extension and refresh the AI page. See the detailed usage guide and changelog.

Constraints and permissions

  • The extension connects only to allowed local loopback addresses and includes no third-party telemetry. Tool results inserted into a third-party AI page can be read by that page, so do not pass credentials or secrets through it.
  • Project file and command permissions can only be changed from the extension popup, using a separate Nexus admin token — not the Bridge Token or any web AI API key. Web content cannot raise permissions.
  • "Project code execution" is not an OS-level sandbox; commands can still affect resources outside the project.
  • The circuit breaker is on by default and caps consecutive tool calls at 20 rounds. Turning it off removes round, duplicate-result and consecutive-failure pauses but does not grant permissions.
  • Automatic sending also sends existing drafts; the extension warns the first time it is enabled.
  • Requires Node.js 22+ to run the test suite (npm test). The original code is licensed GPL-3.0-only.

HotRepo's take

In the recorded new-project snapshot from 2026-11-16, WebAgent had 105 stars and 28 forks in the JavaScript new list, with a reported gain of 5 stars and 2 forks over roughly 1.5 hours. That is a short observation window and does not establish a trend. The project's own README describes 136 tests and a separately verified end-to-end Gemini file-creation loop, which gives the project a concrete testing story beyond the extension shell.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. ysunyang4/NexusA local MCP control plane that lets web-based AI assistants inspect, edit, build and run code on a developer's machine within sandboxed project boundaries.TypeScript119
  2. domsvoboda/xodus-nexusA desktop launcher that lets Linux users browse, install and play Microsoft Store PC games they already own.TypeScript941
  3. vercel-labs/agent-browserA Rust CLI that lets AI agents drive a browser command by command, reading compact page snapshots and acting on element refs.Rust50,711