Jexperiment

Legible makes any small business readable to AI agents. AI agents already shop for people, but they often can’t finish the job on a small business’s website. Prices sit inside images, rules are buried in FAQs, and calendars hide behind contact forms, so the agent gives up or guesses. Legible gives each business a small public page that any agent can read and act on, and an agent that keeps it accurate. The Legible agent. An owner pastes in their website, and the agent reads it, including images, and drafts a structured fact sheet. It works out what’s missing and asks the owner in plain questions. Only facts the owner has confirmed ever go public. The agent can change the fact sheet but never the page itself, and nothing an outside agent says can change the facts. The resource. The page is generated from the fact sheet and runs on a Cloudflare Worker, so it needs no AI to serve. Agents get readable text and markdown, plus forms to get a quote, ask a question or request a booking. Quotes include variants, add-ons, fees, tax and service zones, and need no personal details. The owner adds one link line to their existing site, and agents can find the page from there. Updating it in plain English. The owner can type something like “We’re closed for the next 2 days” into a text box on the Resource screen. The agent works out the exact dates and asks the owner to confirm. The change then goes live and ends by itself when the dates pass. Owners can also click any line or price to edit it, and undo through version history. Awareness of what agents do. Questions from agents land in an Inbox and booking requests land in Bookings, where the owner’s reply appears on a public status page. A periodic review turns this activity into demand insights, like “1 agent asked for ceramic coating”, so owners see demand their website could never show them. Errand, our consumer agent. To test all this, we built a general-purpose assistant on Muse Spark 1.3. It runs a tool loop with multi-query search, page reading and form use, and it can delegate to a browser worker. It knows nothing about Legible, and we froze it before any benchmark run. Agent simulation. Errand doubles as the simulator. When an owner runs a check, test customers ask realistic questions against the business’s page. If a run fails because a fact is missing, that gap becomes a new question for the owner, so each answer improves the page. The result. We tested Errand on a fictional detailing business’s website, with and without one link to a Legible page, and that link was the only difference. Correct answers went from 3 of 7 to 7 of 7, and Errand found the page on its own in 6 of the 7 runs. Runs needing a browser fell from 5 of 7 to none, average time fell from 165 to 32 seconds, and average cost fell from $0.263 to $0.071. Legible is a working prototype, built with Muse Spark through OpenRouter, Exa and a plain crawl for reading websites, browser-use and Cloudflare Workers.