TokenHero — save tokens on every page

Cuts what your AI agent spends reading a web page: text, structure and numbers instead of screenshots. Median 11x fewer tokens.

TokenHero makes the page your AI agent is looking at dramatically cheaper for it to understand. Agents "see" a web page by taking screenshots. One screen costs hundreds of tokens, a long page costs thousands, and the agent still cannot read an image address, tell you which links are broken, see a JavaScript error, or measure a contrast ratio. TokenHero gives it the page as text, structure and numbers instead — and tells you exactly what each call costs before you spend it. WHAT IT DOES • Reads the page as clean Markdown, a heading outline, plain text, or every control with its role and state • Extracts tables as JSON or CSV you can hand straight to a spreadsheet • Lists every image with its true pixel size, plus files, media and a per-host link count • Answers narrow questions cheaply: is this page behind a login? what is broken on it? where does this word appear? does anything overlap, clip or fail contrast? • Shows the console — uncaught JavaScript errors, warnings and failed requests. A crash is invisible in a screenshot. • Says which CSS rule set a property, in which file, on which line, and what it overrode • Takes screenshots when pixels really are the answer — cropped to one element, an explicit rectangle, a whole page in one piece, or mobile/tablet/desktop widths in a single call • Captures only the part of the page that changed since you last looked • Downloads every image above a size you choose, straight to your Downloads folder — the agent is told only how many, so the files never enter its context • Checks every link in the background and reports only the dead ones WHAT IT COSTS Every result carries an est_cost field, calibrated to how the major model families actually price images — Claude, GPT-5, GPT-4o and Gemini count pixels differently, and TokenHero knows the difference. You are never guessing. Measured across three real pages — a 242-row Wikipedia table, an Apple Store product page and an MDN API reference — against reading the same page as text: Is this page behind a login. 41 tokens, between 42x and 378x cheaper. Summarise this data table. 8 to 128 tokens, up to 683x cheaper, the largest saving measured. Which network calls failed. 46 to 351 tokens, with the HTTP status of each. What is broken on this page. 39 tokens on MDN, against 1,703 to read it. Which CSS rule set this colour. 120 tokens, naming the rule, its file and its line. Across the narrow questions agents actually ask, the median saving is about 40x. And, honestly: if one screenshot genuinely answers your question, take the screenshot. TokenHero costs more than a single screen for "what is this page about". It wins on narrow questions, and on the ones pixels cannot answer at all. HOW AGENTS USE IT Open the popup and turn on Agent access. That exposes a function on the page which Claude, Codex, or any agent that can run JavaScript in your browser can call directly. It is off until you turn it on, and a second switch decides whether it follows you as you browse. A third — Adaptive, on by default — lets the agent pick the mode and resolution itself, since it can see both the page and its own budget. You can also use it by hand: click a mode in the popup, see the token cost, copy the result. PRIVACY No server, no account, no telemetry. Nothing is sent anywhere. The extension does nothing at all until you click a button or an agent calls the bridge you enabled.