Core Web Vitals code fixes
Exact git diffs for LCP, INP and CLS regressions — image priority, font loading, layout reservation, main-thread yielding.
LCPINPCLS
For the community · Free Website Health Agent
A free AI webmaster for every small website.
Every small business, nonprofit, school and open-source project deserves a full-time webmaster. We're building one on Claude — an agent that audits your site, writes the fixes and keeps it healthy, free for the community.
Year-one goal: 10,000 community sites audited free
- Free forever for individuals
- Open source at launch
- Powered by Claude agents
Try the first check now. 8 live DNS & email-security checks on your domain.
Live result, not a demo: your browser queries Cloudflare's public DNS-over-HTTPS resolver directly. We don't store what you enter.
// what we're building
The Autonomous Webmaster Agent
Today, keeping a site fast, findable and accessible takes a person most sites never have. We're turning that job into an agent: it watches every page and every deploy, triages regressions, writes the code fix or the structured data, opens a pull request, re-verifies after merge and sends a weekly report. Within 24 months we want site health to be something an agent simply keeps right — the way CI keeps tests green. The free health agent is its first stage.
In development · early access
preview / sites / acme-store
Sample · Preview
Claude Agent Console
diffsample patch
Interactive preview with sample data (acme-store.com is fictional). Switch between LCP, CLS and INP to see the kind of reasoning and patch the agent is designed to produce.
// capabilities
An on-call engineer, SEO lead and schema expert in one agent
What we're building the agent to do: find problems, explain them in plain terms, and ship the fix as reviewable code.
Technical SEO, site-wide
Reasons across hundreds of pages at once in Claude's 1M-token context to untangle canonical chains, hreflang conflicts and indexing anomalies.
Structured data (JSON-LD)
Writes Schema.org markup from your real content and validates it against the schema before it ships.
Works with your stack
Multi-step tool use that integrates with GitHub, Cloudflare and headless CMS APIs — scoped permissions, every action logged.
AI search optimization (GEO)
Structures content, entities and citations so AI answer engines — Claude, Perplexity, ChatGPT Search — can quote your pages accurately.
Regression guard
Checks preview deployments and warns before a merge degrades Core Web Vitals, breaks robots rules or drops pages from the sitemap.
// the agent behind it
A multi-step Claude agent, with a human at the end
One audit is not one prompt. It's a loop of tool calls and model calls across every page of a site, repeated on every run. This is the stage that runs on the Claude API.
-
01 triggerno model
A run starts
A deploy webhook, a weekly schedule, or a new URL in the sitemap kicks off a run.
-
02 crawl & classifyHaiku 5.5
Read every page
Renders each page and classifies it — template, indexability, what changed. Flags broken links, missing alt text and heading problems.
fetch_page · render_headless · read_sitemap
-
03 measuretools
Collect real numbers
Lighthouse lab runs and Chrome UX Report field data for every template that changed.
run_lighthouse · query_crux
-
04 diagnoseSonnet 5.5
Find the root cause
Reads the trace and the repository, then maps each problem to the component and line that caused it.
read_file · search_repo
-
05 fix & verifySonnet 5.5
Write it, then check it
Writes a minimal diff or JSON-LD block on a branch, re-measures on the preview deploy and reviews its own change against the before/after numbers.
propose_patch · validate_jsonld · run_lighthouse
-
06 human reviewa person decides
Pull request or fix list
Opens a pull request — or sends a plain-English fix list — with metrics and reasoning. A person merges; the agent re-verifies after merge and writes the weekly report.
open_pull_request · write_report
Which model does what
- Sonnet 5.5
- Planner — decides which pages and checks a run needs.
- Sonnet 5.5
- Diagnoser & writer — root cause, framework-aware diffs, JSON-LD, plain-English fix lists.
- Sonnet 5.5
- Reviewer — checks its own patch against before/after metrics; writes the weekly report.
- Haiku 5.5
- Classifier & auditor — high-volume page classification, link and accessibility triage, log triage.
Guardrails
- Crawls only public pages, honors robots.txt and backs off on 429/5xx.
- For GitHub-hosted sites, a GitHub App with contents and pull-request access on the repos you choose — its own branches only.
- Never force-pushes, never merges. Branch protection and human review stay in control.
- Every model output goes through strict tool definitions and parses against a JSON schema.
Built on Anthropic Claude
Why we build on Claude
Claude Sonnet 5.5 is the reasoning core for diagnosis, code and writing. Claude Haiku 5.5 handles high-volume page classification and triage during crawls.
Code reasoning that passes review
Frontend fixes need framework awareness — React Server Components, hydration, bundler config. Claude writes minimal, idiomatic diffs.
Reliable structured output
Findings, schema markup and fix plans come back through strict tool definitions, so every response parses against our JSON schema.
Prompt caching for repeated crawls
Site maps, design-system rules and repo context are cached once and reused across page analyses, so a free tier stays affordable to run.
agent/diagnose.ts
sketch · tool use · cached
// Repo context is cached; each page adds only its trace
const msg = await client.messages.create({
model: "claude-sonnet-5-5",
max_tokens: 16000,
system: [{ type: "text", text: repoContext,
cache_control: { type: "ephemeral" } }],
tools: [readFile, proposePatch, runLighthouse],
messages: [{ role: "user",
content: pageTrace }],
});1M
token context
Sonnet 5.5
diagnosis & diffs
Haiku 5.5
crawl triage
// api usage
Why we need Claude API credits
The free health agent is not a rules engine with an AI label. Deciding why a page is slow, which broken link matters, what structured data a page should carry — and explaining it so a bakery owner or a school volunteer can act on it — is language-model work, on every page, on every run. Without an LLM API this stage doesn't exist.
It also never finishes: sites change, so the audit repeats weekly for community sites (and on every deploy for paid sites). Usage grows with every site we take on — which is the point of a free tier for the whole community.
Credits go to the free community tier first.
| Task | Model | Tokens per run | Runs / month | Monthly tokens |
|---|---|---|---|---|
| Classify & audit one page | Haiku 5.5 | ~2.5k in + ~0.3k out ≈ 2.8k | 50 pages × 4 × 500 = 100,000 | ≈ 280M |
| Diagnose one site | Sonnet 5.5 | ~40k in + ~4k out ≈ 44k | 4 × 500 = 2,000 | ≈ 88M |
| Plain-English fix list + JSON-LD | Sonnet 5.5 | ~12k in + ~3k out ≈ 15k | 4 × 500 = 2,000 | ≈ 30M |
| Pull request (GitHub-hosted sites) | Sonnet 5.5 | ~60k in + ~6k out ≈ 66k | 125 sites × 2 = 250 | ≈ 16.5M |
| Launch cohort total: 500 sites (≈ 829k tokens per site per month) | ≈ 414.5M / month | |||
| 12-month target: 1,200 sites on weekly audits × 829k | ≈ 994.8M (~1B) / month | |||
Per site per month: 560k (pages) + 176k (diagnosis) + 60k (fix list) + 33k (pull requests, averaged over the 1 in 4 sites hosted on GitHub) = 829k tokens. One-off first audits on the way to the 10,000-site goal come on top. Prompt caching reuses each site's map and repo context across pages; that lowers cost, but every page still needs its own model call. The paid Autonomous Webmaster Agent runs the same pipeline on every deploy and is meant to pay for itself through subscriptions.
// roadmap
Where we're going
A 24-month plan. Everything below is a goal, not a result.
-
now → month 3 · building
Free Website Health Agent
- Live DNS & email-security check (on this page today)
- Crawler + Claude audit pipeline
- Launch cohort: 500 community sites on weekly audits
-
months 3–12
Open source, at community scale
- Public repository and MIT license at launch
- Pull requests for GitHub-hosted community sites
- Goal: 10,000 sites audited free, 1,200 on weekly audits
-
months 12–24
A full-time AI webmaster for every site
- Autonomous Webmaster Agent (paid): every page, every deploy
- Regression triage, PRs, re-verify after merge, weekly report
- Paid teams fund the free tier for the community
// faq
FAQ
Is this a finished product?
Not yet — we're early and building fast. The DNS check at the top of this page is live today; the full agent is in development and we're opening early access in waves. Anything marked Sample or Preview is illustrative, and numbers on this page are targets.
Is the community tier really free?
Yes — free forever for individuals, nonprofits, schools and open-source projects. Small businesses can use the free tier too. The paid agent is for teams that want every deploy monitored and fixed automatically. Details.
Will crawling slow down my site?
No. The crawler honors robots.txt and crawl-delay, adapts its rate to your server's response times and backs off automatically on 429/5xx.
Is my content or code used to train AI models?
No. We use Anthropic's commercial API, which does not train on customer inputs or outputs by default. Repository access is limited to what a fix requires, and code context is not kept after the job.
Every small website deserves a webmaster.
Join early access for the free health agent, or talk to us about the Autonomous Webmaster Agent for your team.