Ai agent Blue
A Manifest V3 Chrome extension that operates your browser from a plain-English goal. You type what you want in a side panel; the AI agent plans the task, then reads each page, clicks, types and moves between tabs until it is done.
Overview
Most knowledge work is repetitive multi-step web tasks: research across a dozen tabs, copy data between sites, triage email, compile notes into a document. Ai agent Blue does that work for you, inside your own browser.
Unlike cloud "AI browser" tools, it runs entirely on your machine with your own API key. Your logged-in sessions stay yours, and an on-device mode using Chrome's built-in Gemini Nano sends nothing off your computer. It is free to run on standard free API tiers, with no backend server.
Install (load unpacked)
- Download and unzip
ai-agent-blue-v3.1.zip. - Open
chrome://extensionsand turn on Developer mode (top right). - Click Load unpacked and select the unzipped
ai-agent-bluefolder. - Click the extension icon to open the side panel.
- Get a free API key: Google AI Studio ↗ (the panel also links to it). Paste it, or pick Chrome Built-in AI to run with no key at all.
- Leave the model on the suggested
gemini-3.1-flash-lite(free tier, fast). Type a goal, hit Start Agent.
Requires Chrome 138+. The Gemini free tier is enough to run tasks; on-device mode (Chrome Built-in AI) needs no key or connection at all.
What it can do
Plans the task
One reasoning pass turns your goal into a checklist and the data slots it needs, shown live and checked off as it goes.
Works across tabs
Opens, reads, and switches between tabs. Carries gathered data in named memory slots and injects it into other fields intact.
Asks before it acts
Pauses for your approval before send, post, buy, delete, or closing your tabs. One toggle for full autonomy.
Injection-safe
Treats all page text as data, never instructions. A hostile page cannot redirect the agent.
Runs on a small model
Engineered so a free, lightweight model stays reliable: stuck-loop detection, JSON repair, crash recovery.
Voice + repeat
Dictate a goal by voice, or run the same task N times in a row.
Example goals
"Open three pricing pages, pull each tool's free plan and cheapest paid price, and put it all in a new Google Doc as a table."
"Read the README of this GitHub repo and write a one-page explainer in a Google Doc."
"Draft a reply to my newest email, but stop and ask me before sending."
"Close every tab except this one and my pinned tabs."
Under the hood
Side panel + service worker + a sense → think → act loop. The content script serialises each page into a numbered list of elements; the model gets that list plus the goal, plan and a compact memory view, and returns one JSON action per step. No vision model, no screenshots, so it runs on any text LLM.
Full write-up: ARCHITECTURE.md in the repo.