OpenSider: Let the Browser Drive Your Local AI Agent

OpenSider is a free, open-source browser extension that integrates your local CLI agents into the browser sidebar, so you can work with an agent directly inside the browser window you already use.

Repo (installs, tries and stars are all appreciated): https://github.com/parksben/opensider

A while ago I was researching several online products, comparing tech stacks, service pricing and licensing terms across multiple dimensions. That information is scattered across many web pages — you have to read them one by one and then compile it into a table you can compare side by side.

This kind of work is best handed to a local AI agent. I usually use Claude Code and Codex; writing code and scripts with them is smooth.

What actually blocked me was something small: the agent can't see the dozen or so tabs I have open in my browser, and it has no idea what my actual browsing/interaction state is on those pages.

So I tried three approaches — each with significant shortcomings:

Approach	Reuses your real browser	Chat entry point	Cross-page operations
Playwright / CDP	No. Starts a clean browser; login state and forms don't carry over	Terminal or scripts	You orchestrate the flow yourself
Vendor-official plugins	Yes	Editor sidebar	Limited by the plugin's capabilities, and locked to the vendor's model
Browser MCP	Yes	Terminal	You have to issue an instruction for every page you open

All three can work, but none are smooth. Ultimately they all stop at the same problem: what the agent lacks isn't intelligence — it's the ability to perceive and restore the user's browsing context.

So OpenSider was born

Once I thought it through, I inverted the relationship: instead of making the agent reach for the browser, I let the browser hand the current page to the agent.

The idea is to make the browser the client and the local agent the executor. The browser actively pushes "what the user is looking at" to the agent, and the agent uses that information to decide what to do next. To let the agent operate web pages smoothly, the browser also has to expose a set of tools to it.

Based on this idea — and with a lot of help from AI — I shipped an open-source browser extension: OpenSider.

With this extension, a lot of work that used to mean bouncing back and forth between web pages and an agent UI becomes much simpler: you just give the agent instructions right in the browser, and it works alongside you in the same browser context. The agent's ability to operate web pages improves dramatically too — through the APIs the extension exposes, the agent can jump and switch among multiple already-loaded pages to complete a task, without cold-starting a browser window or tab on every tool call.

The following five sections describe the main parts of this extension.

1. Product form

After you install and enable the extension, you can pick a locally installed agent (CLI) in the browser sidebar and chat with it. Any agent that supports the ACP (Agent Client Protocol) can connect — for example Claude Code CLI, GitHub Copilot CLI, OpenCode CLI, and more.

The extension is essentially a connection layer: a ChatUI lets you talk directly to a local agent, while the browser state is synced into the conversation context in real time, so the agent and the user see "the same set of pages".

2. How to use it

Using the extension takes just three steps:

1. Install the extension (send an install prompt to the agent you're using);
2. Launch the extension to connect to your local agent;
3. Chat with the agent directly in the ChatUI.

Taking that product research as an example again, I type this in the browser sidebar:

"Go through the pricing pages of products A/B/C and produce a side-by-side comparison table."

The agent automatically switches to the right tabs, reads the pricing and plan details, generates a comparison table and returns the result. The whole process involves no copying links, no switching to a terminal, and no logging in again.

3. Key features

• Reuse your real browser state: the site's user identity (login state) and the forms you've already filled in are preserved and reused — no need to reconstruct the context.
• Collaborate in place: whatever page you're on, the agent works with that page's information. You can also pick page elements to ask questions about them.
• Web automation: page navigation, reading, clicking, form filling, screenshots and more are injected into the agent as tools, and the automation is shown live on the page (simulated mouse clicks, keyboard input, etc.).
• Freely switch agents and models: switch agent or model with one click, and use different agents within the same session — no lock-in to a single terminal tool.
• Data and artifacts stay local: session data and agent artifacts are stored locally and can be reused.

4. How it works

Core approach: the browser and the local agent are bridged by a Chrome Native Messaging Host. It starts when the sidebar connects and exits when it disconnects — the process is not resident. This host launches the CLI you choose over the ACP protocol and exposes the web-automation tools to the agent through the same connection. No extra background service is required.

Communication between the extension and the page doesn't define a new protocol either — it's based directly on file I/O. A full interaction cycle looks like this:

• The extension writes browser tab information into several files under the workspace;
• the agent reads these files in real time to learn the page content and the available controls;
• when the agent needs to operate the page, it writes instructions into the commands/ directory;
• the extension executes the instructions and writes the results back to results/.

To an agent, operating a web page is no different in essence from reading and writing a file.

5. Installing OpenSider

To accommodate different platforms (macOS / Windows / Linux, etc.) and specific runtime environments, OpenSider doesn't ship the installer as a fixed release artifact. Instead it provides an online Skill plus a set of tool scripts (including binaries) for the agent.

Just send an install prompt to your usual agent app, and it will walk you through installing the extension:

Help me install the OpenSider browser extension.
First read https://raw.githubusercontent.com/parksben/opensider/main/skills/opensider/SKILL.md and follow its installation process.

More information

For full details about the project, head over to the GitHub repo: https://github.com/parksben/opensider

You're welcome to install it, try it, and give feedback to help me keep improving the tool. A star is a huge support for the project — thank you!