Reading the web with my own AI: the extension, a month in

I read a lot on the web — long articles, documentation, research write-ups, the occasional rabbit hole. The friction was always the same: when I wanted to do something with what I was reading — summarize it, ask a question, double-check a claim — I had to copy text out of the page, paste it into some other tab, and lose my place. So a couple of weeks ago I shipped the smallest possible fix: a browser extension that docks a side panel next to the page and summarizes or answers questions about it, streamed from my own inference.club cluster — my models, my data, nothing in the middle.

That first release was deliberately small: two surfaces, two actions. Then I kept going. This is a tour of what it has become — what it does now, how it works, and where I want to take it next. It's still open source and it's still a thin client of the /v1 API the playground already uses; the server hasn't gained a single line of code on its behalf.

What it is

The extension is a thin, open client for an OpenAI-compatible inference endpoint — in my case my inference.club account, but it points at anything that speaks the same API, including a model running on localhost. You paste an API token once, and from then on every article you open gets a little AI workbench docked to its side.

The important framing hasn't changed since the first post, and it's still the whole point: it's a copilot for the page in front of you, not a data funnel. The page is read locally in your browser, nothing is sent anywhere until you click an action, and when it is sent it goes only to the endpoint you configured. No analytics, no third parties, no servers of mine in the middle.

How it works

Under the hood it's deliberately simple, which is part of why it's pleasant to use:

• Local extraction. When the panel is open it reads the readable version of the current tab using Mozilla's Readability — the same engine behind Firefox's Reader Mode. You get the article text, title, byline, and metadata, with the page's nav, ads, and chrome stripped away. This happens entirely on your machine.
• Streaming chat. Actions turn that text into a prompt and stream the response back token-by-token over the OpenAI-compatible /v1/chat/completions endpoint. You watch the answer appear instead of staring at a spinner.
• Your endpoint, your rules. The base URL and token are yours. Swap in a different model, or point it at a self-hosted backend, and nothing else changes.
• Built to stay out of the way. It's a Manifest V3 extension (built with WXT + Vue 3 + TypeScript + Tailwind) whose primary surface is the Chrome side panel, so the AI sits beside what you're reading rather than covering it.

Everything you generate stays on your device: preferences and chat history live in local browser storage, never synced off the machine.

The features

Summarize, your way

One click turns the page into a TL;DR, bullet points, key takeaways, or a plain-language ELI5 explainer. Great for triaging whether a long piece is worth your full attention.

Ask — a real conversation about the page

Ask follow-up questions and get answers grounded in the article. It's multi-turn: "summarize this," then "expand on point 3," then "is that claim actually supported?" — it remembers the thread. Answers render as proper Markdown, so headings, lists, tables, and code come out formatted instead of as a wall of text.

Attach images from the page

Sometimes the interesting part of a page is a chart, a diagram, or a screenshot. A ＋ button scans the current page for images and lets you pick a few to attach to your question — they're sent to the model as vision input. "What does this graph show?" finally has an obvious answer.

Speed reader

This is my favorite. Any article or any AI response can be sent to an immersive speed reader that flashes one word at a time (RSVP), with an optimal-recognition-point highlight to keep your eyes still. Play, pause, resume, scrub through the text, and tune the words-per-minute, font, size, color, and how many context words to show around the focus word. It turns "I'll read this later" into "I'll read this now, in two minutes."

Persistent history, organized by site

Every conversation is saved per page URL and browsable in a History view, grouped by website and filterable by text. Jump back to a past thread, or click straight through to the original source page. Your reading and your questions become a little searchable knowledge trail instead of disappearing when you close the tab.

It follows you as you browse

Switch tabs or navigate to a new article and the panel automatically re-reads the page in front of you. No re-clicking, no stale content — the copilot is always about whatever you're actually looking at.

Themes that are easy on the eyes

Ten hand-tuned light and dark palettes — Daylight, Paper, Mist, Midnight, Nord, Dracula, Solarized, Gruvbox, Forest, Rosé — plus "match system." The theme picker shows a live preview of each one, so you can pick something comfortable for a long reading session in dim light.

Advanced mode for the curious

Flip on Advanced mode and each response reveals what happened under the hood: time to first token, total time, token usage, finish reason, and request size — plus full page metadata. You can even expand the exact system prompt and article context that was sent to the model, so there's no mystery about what the AI was actually given.

Streaming toggle

Streaming is on by default, but you can switch to "show the whole reply at once" if you prefer a cleaner, all-at-once result.

How it actually changes my browsing

A few habits have genuinely shifted:

• I triage before I commit. A quick bullet summary tells me whether a 4,000-word piece deserves my next twenty minutes.
• I interrogate what I read. Instead of passively trusting an article, I ask it to check claims, define jargon, or compare two sections — right there, in context.
• I read faster when I want to. The speed reader gets me through a backlog of "save for later" tabs that I'd otherwise never reopen.
• My reading leaves a trail. Because every thread is saved by URL, I can come back weeks later and find both the source and what I asked about it.
• It's mine. Running on my own cluster means I'm not feeding a third party my reading habits, and I can use whatever model I want.

The throughline is that the web stops being a one-way stream of text and becomes something I can converse with, on my terms, without leaving the page.

Next steps

Plenty I want to build from here:

• Narration / text-to-speech. Listen to an article or its summary read aloud, with the speed reader's word highlighting synced to the audio — reading and listening at once, perfect for walks or tired eyes.
• Image generation. Generate diagrams, hero images, or "explain-this-as-a-picture" visuals from the page or a summary, and drop them right into the thread.
• Research mode. Let the model go beyond the single page — web search plus synthesis across multiple sources via an agent loop, with citations, so a question can pull in context the article didn't have.
• Selection-aware actions. Highlight a paragraph and act on just that — "explain this," "translate this," "steelman this argument."
• Smarter image handling. Auto-downscale attached images to keep payloads small, and let me draw a box on the page to capture a region instead of picking whole images.
• Keyboard-first flow. Shortcuts to open the panel, summarize, start the speed reader, and send a follow-up — never touch the mouse.
• Export and share. Save a thread (and its source link) to Markdown, or push it to a notes app, so insights flow out into the rest of my tools.
• Search across my history. Semantic search over every conversation I've had — "what was that article about retrieval caching, and what did I ask about it?"
• More browsers and surfaces. A Firefox build, and maybe a mobile companion.

Several of those — narration, image generation, research — already have endpoints my cluster serves today; the extension's job is just to point them at the page in front of me. The fun part is that the foundation — local extraction, your own model, a calm side panel — makes all of these feel like natural extensions rather than bolt-ons. The web is the input; the copilot is whatever I decide to make of it next.

If you want to try it: the extension is open source at inference-club/browser-extension — clone it and run it unpacked today. You'll need an inference.club account and your API token, and a cluster to point it at. If you don't have hardware on the network yet, the run-an-agent guide gets your GPU serving in a few minutes, and then your reading copilot runs on it.