Agent Ready
Lighthouse for AI agents. People increasingly ask Claude, ChatGPT and Perplexity to find a restaurant, book a service or compare prices, and the agent too often comes back with "sorry, I couldn't find the menu." Agent Ready scores how well AI agents can use a website and tells the owner, in plain language, what to fix.
Live: https://agentready.sfo.unikraft.app
How it works
- Checks the page. 60 checks read the real site: can agents get in (robots.txt, bot walls), understand it (hours, location, menus, prices, contact) and take action (MCP/A2A/OpenAPI endpoints, declared actions, labeled forms)?
- Sends an agent. A Claude agent (Sonnet, run through Omnara) tries the things a customer would ask for, like finding prices or contact details, and records every step.
- Gives one score. Checks and agent results combine into a single 0–100 score and grade across six weighted categories, with fixes ranked by impact for that kind of business.
- Hands over a fix link. The owner pastes one link into their site builder's AI (Wix, Squarespace, Lovable, Bolt, Claude) or sends it to their developer, then re-checks.
Sites that block bots get no score. Instead the report names the blocker (Cloudflare, DataDome, Akamai and others), explains how to let agents in, and lists the site in its own section of the leaderboard.
More on the problem and who it's for: docs/problem-solution.md. How the single score is produced: docs/one-final-score.md.
Use it
- Web:
/report?url=example.com
- Leaderboard:
/leaderboard, filter by business type with ?type=restaurant
- JSON API:
GET /api/audit?url=example.com or POST /api/audit with { "url": "..." } (the page checklist only, without the AI agent, and not added to the leaderboard)
- MCP server:
/api/mcp (streamable HTTP), tools audit_site and generate_fixes
claude mcp add --transport http agent-ready https://agentready.sfo.unikraft.app/api/mcp
Built with
Next.js 15 and TypeScript, Supabase (audits, leaderboard, fix links), Claude Sonnet via Omnara (the agent), Context.dev (logos and business categories), Stripe test mode (the $9 fix link, see docs/payments.md), and Unikraft Cloud (hosting), deployed by GitHub Actions.
Run it locally
pnpm install
cp .env.example .env # fill in the keys you have
pnpm dev # http://localhost:3000
Every key is optional; the app degrades instead of failing. Without Supabase nothing is saved, without the Omnara vars the agent step fails, without Context.dev there are no logos and categories fall back to keywords, and without Stripe fix links are free. For Supabase, run supabase/schema.sql and src/lib/agent/schema.sql in your project's SQL editor.
Code map
- Checks:
src/lib/audit/checks.ts, one entry per line of evals/AGENT_CHECKS.md. Each is a { id, category, title, weight, method, run(ctx) } object, so adding one is a single array entry.
- Scoring:
src/lib/audit/rubric.ts. Checks are rated 0–4, na or unknown (unknown lowers coverage, never the score). Categories A–F are weighted 15/20/25/15/15/10, and a category counts once 25% of it is measured.
- Blocked sites:
src/lib/audit/blocked.ts detects the bot wall and writes the unblock steps.
- Agent:
src/lib/agent/. With AGENT_RUNTIME=omnara, Omnara runs the loop and model and our tools (src/lib/agent/tools.ts) run as custom tools on our server; otherwise src/lib/agent/model.ts calls Claude directly (ANTHROPIC_API_KEY). Guardrails are in src/lib/agent/guard.ts.
- Pages and API:
src/app/ (report, leaderboard, /api/agent, /api/audit, /api/mcp).
pnpm test && npx tsc --noEmit && pnpm lint # what CI runs
AGENT_STUB=1 pnpm agent https://example.com "find pricing" # agent with fake steps, no model
pnpm agent https://example.com "book a demo" # real agent run
Deploy
Every push to main deploys to Unikraft Cloud (.github/workflows/deploy.yml): it runs the tests, builds the image, swaps it onto the running instance and checks the site. Pull requests get tests and a production build first (ci.yml).
The workflow needs the UNIKRAFT_TOKEN secret and the UNIKRAFT_ORG, UNIKRAFT_INSTANCE and UNIKRAFT_SERVICE variables (the service name sets the URL, https://<service>.sfo.unikraft.app). If the instance doesn't exist, the first run creates it.
The instance's env vars come from repository secrets named as in .env.example, and every deploy sets them again. To rotate a key, update the secret (gh secret set CONTEXT_DEV_API_KEY) and re-run the workflow from the Actions tab. The optional AGENT_DAILY_BUDGET variable sets how many new agent runs a day are allowed (default 50).
