AgentReady Sites methodology
This directory helps an agent discover websites by task and understand how each site can be accessed. A site card reports observations from a dated scan.
What a card means
- Tasks are normalized directory labels mapped from existing catalog categories. They describe a likely use, not a demonstrated capability of the site.
- Access level describes the path we could identify: keyless URL, open MCP, MCP with signup, API with signup, or browser only.
- Discovery signals show which public agent-facing files or links were found and when they were checked.
- An MCP endpoint is marked open only after initialize and tools/list succeed. We store the check date and tool count, not returned tool names.
- A missing or untested signal is not evidence that a site is unusable.
What this directory does not claim
Discovery signals and task evaluations are separate. A task result applies only to the published task, agent/model/browser configuration, date and sample shown. It does not certify usability, safety, security, accessibility, quality or endorsement.
Scanning
The scanner identifies itself, uses HTTPS, checks robots.txt before other paths, limits requests per domain, applies a timeout, and caches results for a week. The first-party source URL and check date are shown when a signal is found. The scanner does not perform purchases, publish content, send messages, or use private accounts.
Everyday-sites pilot cohort
Besides the websites of vendors in the AgentReady catalog, the directory includes a small pilot cohort of everyday public websites (26 sites, selected 2026-11-21, cohort id everyday-2026-11). Its cards carry the label “Everyday-sites pilot cohort”. The list with each site’s rationale and source is in data/everyday-sites.json.
How it was chosen. Hand-picked by AgentReady on 2026-11-21 to cover the everyday tasks in data/tasks.json that people ask assistants about (reference, news, weather, travel, flights, local, shopping, jobs, video, government, health, parcels). Each site is public and readable without an account. Excluded: sites that need a login for the read-only scan, adult or gambling sites, and domains already in the directory. This is a representative pilot, not a ranking: no traffic ranking was used, and inclusion does not mean a site is AgentReady or endorsed.
Limitations. About two dozen sites cannot represent the web. The choice leans to English-language and a few Israeli public services. Discovery signals describe files a site publishes, not whether an agent can finish a task there; task results come only from separate evaluations.
Results for this cohort are reported separately from the vendor seed.
Corrections
Site owners can request a factual correction or ask for a rescan through the AgentReady Tools contact page. Independently observed scan results remain attributed to the scan and its date.
Task evaluations
Each evaluated task has an observable success criterion and is repeated across identified agent/browser configurations. We publish successes, failures and blocked runs, sample size, last-tested date, tool calls or interaction steps, retries, access friction and a public evidence source when available. Different tasks and configurations remain separate.
We do not calculate an overall site score or rank sites. A small sample is descriptive and can change when a site, prompt, model, browser or network changes. We publish no synthetic fixture data as results. Untested sites are labelled “Not evaluated.”
Pilot task safety
Pilot tasks use public information and read-only actions. We do not create accounts, submit forms, send messages, purchase, delete or publish content on real accounts. Test evidence must not include credentials or personal data. We identify the tested configuration and stop when a site asks for authentication, blocks automation, or presents a rate limit.