Frequently asked questions

AgentReady is an agent interface platform. Paste any public URL and AgentReady crawls the site, generates a spec-compliant llms.txt, and hosts a live MCP server and cited RAG /ask endpoint — no code required. Any MCP-compatible AI client (Claude Desktop, Cursor, Windsurf, VS Code Copilot) can then query the site with cited, multi-page answers.

AI agents that use web_fetch to read websites face three systematic failures: (1) JavaScript-rendered sites (React, Next.js, Vue) return empty HTML — agents get no content. (2) Answers that span multiple pages require multiple fetches, but agents typically fetch one URL. (3) web_fetch returns raw text with zero citations. AgentReady fixes all three: a 4-layer crawl pipeline handles JS rendering, the vector index retrieves across all pages, and every answer includes source URLs.

A website is agent-native when AI agents can reliably discover it, query it with natural language, get cited answers, and understand what actions the site supports. Most websites are browser-native — designed for humans clicking and scrolling. AgentReady adds the agent-native layer on top of any existing website without requiring changes to the original site.

Two audiences: (1) AI agent developers who need reliable, cited answers from any website — use the ask_site MCP tool or the /ask REST endpoint. (2) SaaS and documentation site owners who want their site to be queryable by every AI agent — index your site so agents answer about you correctly instead of hallucinating.

AgentReady uses a 4-layer pipeline: (1) llms.txt / llms-full.txt — if the site publishes one, it is used as a high-quality structured source. (2) Standard HTML crawl — fetches pages via sitemap or link crawling, extracts clean text. (3) __NEXT_DATA__ extraction — for Next.js apps, parses server-embedded data. (4) Jina Reader fallback — for JS-only SPAs that return empty HTML, uses a remote rendering service to extract content. Indexing takes about 60 seconds for a typical site.

AgentReady hosts a global MCP server at /api/mcp with 8 tools: ask_site, list_sites, get_site_capabilities, get_site_facts, plan_site_action, submit_site, refresh_site, and rate_answer. Each indexed site also gets its own MCP server at /api/sites/[siteId]/mcp. Connect any MCP-compatible client by URL — no install required.

Every indexed site gets a REST endpoint at /api/sites/[siteId]/ask. POST {"query": "your question"} to get a cited answer. The response includes the answer text, source page URLs, a confidence score, and an HMAC signature you can verify. This works without MCP — useful for scripts, CI pipelines, and non-MCP integrations.

When you ask a question, AgentReady embeds the query (OpenAI text-embedding-3-small), retrieves the top relevant chunks from the site's vector index (pgvector on Supabase), and passes them to Claude Sonnet 4.6 to synthesize an answer. The model is instructed to cite each claim with its source URL. The answer includes both the text and the set of source pages it drew from.

Yes. The 4-layer pipeline includes a Jina Reader fallback that renders JavaScript and extracts content from the DOM. In the benchmark, ask_site achieved 76% fact coverage on JS-heavy sites (Notion, Figma, Webflow, Linear) vs web_fetch's 56%.

Freshness depends on your plan. Free: 5 manual refreshes/month. Pro: 50 refreshes/month, plus you can trigger a refresh from your docs deploy pipeline using the webhook at /api/webhook/refresh or npx @agentreadyweb/mcp refresh yourdomain.com. Team: unlimited refreshes. The site changelog at /api/sites/[id]/changes shows what changed in each crawl.

Yes. The Free plan includes 1 site, 100 pages crawled, 5 index refreshes/month, the public MCP endpoint, llms.txt generation, cited RAG answers, an agent-readiness grade, 7-day analytics, and query intent analytics. No credit card required.

Pro ($49/month) adds: 5 sites, 500 pages per site, 50 refreshes/month, 30-day analytics, private MCP with API key auth, GitHub check runs on every refresh, Slack regression alerts, Agent Simulator, query gap detection with AI-drafted content briefs, answer tests with expected-fact CI, Content Push API, query webhooks, fact attestation, full crawl changelog, and site freshness alerts.

Team ($149/month) includes unlimited sites, unlimited pages, unlimited refreshes, everything in Pro, 90-day analytics, and early feature access.

Yes. Enterprise includes SSO, private deployment, SLA, audit exports, and custom pricing. Contact us at agentready.it.com/enterprise.

Add this to your Claude Desktop config file (~/.claude/claude_desktop_config.json): { "mcpServers": { "agentready": { "command": "npx", "args": ["-y", "@agentreadyweb/mcp"] } } }. Restart Claude Desktop. You now have 8 tools available.

Add to ~/.cursor/mcp.json: { "agentready": { "command": "npx", "args": ["-y", "@agentreadyweb/mcp"] } }. Restart Cursor.

Yes. If your MCP client supports HTTP transport (WebMCP), connect directly by URL: https://www.agentready.it.com/api/mcp. No npm, no install, no account required.

Add a refresh step to your CI/CD pipeline: npx @agentreadyweb/mcp refresh yourdomain.com. Or use the webhook: curl -X POST https://www.agentready.it.com/api/webhook/refresh -d '{"domain": "yourdomain.com"}'. Rate limited to once per hour per domain.

Any public website with accessible HTML content. This includes static sites, server-rendered apps (Next.js, Nuxt, SvelteKit), and JavaScript SPAs via the Jina Reader fallback. Sites behind authentication, CAPTCHA, or that block all crawlers cannot be indexed.

Benchmark (60 cases, 43 sites, GPT-4o judge, July 2026): On JS-heavy sites, ask_site achieves 76% fact coverage vs web_fetch 56% (+20pp). Citation rate: ask_site 100% (60/60) vs web_fetch 0%. High-hallucination answers: ask_site 1 vs web_fetch 6. web_fetch wins on static docs (86% vs 81%) and API reference pages (91% vs 76%). web_fetch is also faster (4.2s vs 7.8s p50 latency).

web_fetch is better for: dense single-page API references (the full spec is on one URL), freshly published content that has not been indexed yet, and when latency matters more than citation accuracy. ask_site is better for: JS-rendered sites, questions that span multiple pages, any time citations are required, and when hallucination risk matters.