What happens when LangGraph.js runs directly inside the browser?
I used to mostly work with the Python side of LangChain/LangGraph.
Then I started experimenting with LangGraph.js directly in the browser while exploring WebMCP, and I wanted to see what it would look like to wire WebMCP tools into a LangGraph.js agent flow.
That slowly turned into Brow: a WIP open-source Chrome side-panel agent that runs in the real browser session.
The goal was to see how far I could push an agent that runs client-side, close to the page, instead of relying only on a backend or an external automation layer.
Brow can already:
- work with both closed frontier models and local/open-source models, using Claude/OpenAI providers or OpenAI-compatible endpoints with custom base URLs
- chat with an agent directly in the Chrome side panel
- run the agent flow client-side in the browser using LangGraph.js
- use the current page and browser context
- discover WebMCP tools exposed by websites
- wire WebMCP tools into the LangGraph.js agent flow
- connect to remote MCP servers
- render MCP Apps directly inside the chat
- use browser automation tools like click, type, scroll, tabs, screenshots, etc.
- record workflows and show them to Brow as reusable context
- use reusable skills to help the agent adapt to specific tasks and websites
For this kind of project, using LangGraph.js directly in the browser is interesting because the agent can live much closer to the actual page: page context, browser tools, WebMCP tools, MCP servers, and UI rendering can all be connected from the extension runtime.
This is still experimental, imperfect, and very much a work in progress.
It started as a side project, built in the quiet hours after work and family time, one tired-but-curious commit at a time.
Small note about the video: it goes a bit fast in some parts, so don’t hesitate to pause. Video editing is definitely not my area of expertise, I mostly wanted to show the current state of the project as clearly as I could.
I’d love to get feedback from people using LangChain or LangGraph, especially on browser agents, client-side agent orchestration, WebMCP/MCP integration, and what kind of use cases this could unlock.
And if anyone is interested in this direction, contributions are very welcome. I’d love to find motivated people who see potential in this and want to help shape it into something bigger than a solo side project.