- Entry date
- 10 March 2025
- Category
- Tooling
- Lead over the world
- a working baseline, not a lead
- Access
- 🔓 Public
The AI didn't just answer questions in the browser — it clicked, filled, and navigated inside one, on real pages, under real permission constraints.
- Manifest V3
- Chrome Extension architecture
- 3
- core capabilities: extract, fill, navigate
Honest evaluation
The narrow claim — an extension can act on a page, not just answer about it, under real Manifest V3 constraints — is demonstrated with a working three-capability tool.
What would prove or disprove it further
What would narrow or disprove it: if the same extract/fill/navigate capabilities failed to hold up against a broader sample of real-world sites (heavier JS frameworks, more aggressive anti-automation measures) than whatever was actually tested here. The report doesn't state the breadth of sites exercised, which is the one gap worth flagging against the verdict.
The evidence — full reasoning behind the verdict
Verdict: proven. The question this project actually asked was narrow: can a browser extension act on a live page — extract, fill, navigate — under the real constraints of Manifest V3, rather than just describe what's on it? The answer is yes, and it shipped as a working, complete extension, not a demo video.
The evidence is the constraint-finding itself: permissions, DOM access, and page-load timing turned out to be the actual bottleneck, not the AI's reasoning — and that finding could only be produced by building against a real browser, which this did. The honest scope limit — a single-purpose tool, not a generalising platform — was never the hypothesis being tested here, so it doesn't count against the verdict; it's an openEdge, not a miss.
Most AI browser tools in early 2025 answered questions about a page. This one acted on the page — which turns out to be a materially harder engineering problem than it sounds.
What it was
ChromeBot — a Chrome/Edge browser extension that lets an AI extract information from a page, fill in forms, and navigate autonomously, built on the Chrome Extension Manifest V3 architecture and an OpenAI API backend.
What we built
ResearchA working extension with three core capabilities — read page content, fill form fields, and move between pages — driven by an LLM deciding what action to take next. Built against Manifest V3's permission model, which is deliberately more restrictive than earlier extension architectures, and against the real unpredictability of live DOM structure and page-load timing.
What we learned — including the honest negative(s)
The real education here was architectural, not conceptual: permissions, DOM access, and page-load timing are the actual hard constraints of running AI-driven interaction inside a browser, and none of them show up until you build against them directly. A page that hasn't finished loading, a DOM structure that differs from what the model expects, a permission the extension doesn't have — each of these breaks an otherwise-correct plan in a way that has nothing to do with the AI's reasoning quality.
The honest edge: this is a working single-purpose tool, not a platform. It does one thing — act inside a browser tab — and does not generalise beyond that scope on its own.
Where it went / status
Working and complete as a single-purpose extension. Three months later, the same underlying instinct — an AI acting inside a browser, not just answering about it — scaled up into a considerably more ambitious browser-as-agent platform, carrying forward everything learned here about permissions and page-timing constraints.
The narrower point is worth keeping on the record: an extension that can act on a page, not merely read it, is a meaningfully different engineering problem than a chatbot with browser context. ChromeBot is the first place that distinction got made concrete, in code, against the real constraints of Manifest V3 rather than in the abstract.
What is still open — kept visible
The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.
- A working single-purpose tool, not a platform — it does not generalise beyond page extraction, form-filling, and navigation.
- Real browser constraints (permissions, DOM access, page-load timing) bounded what was reliably achievable, and those limits are the useful finding here, not a footnote.
Where this connects
Sources
- vocabotics project audit — ChromeBot (AI-powered browser interaction extension), Mar 2025vocabotics internal project history · as of March 2025