AI is reading your website
completely wrong.
Onto is the compatibility layer for the agent web. AI developers call our API to read any site cleanly. Site owners install our SDK to serve themselves cleanly. Agents will soon act on the web through the same engine. Same compatibility layer. Three buyer surfaces. One destination.
One API call. From 812 KB of noise to 8 KB of signal.
Use it from your code. Use it from Claude Code, Cursor, or any MCP client.
Drop any URL. Get your AI-readiness report in seconds.
800 KB of HTML. 8 KB of signal.
The modern web was built for human eyes. AI agents reading the same pages parse hundreds of kilobytes of layout code to find a few kilobytes of actual content. They hallucinate. They burn token budgets. They cite the wrong information under your brand name. Onto fixes this at the protocol layer.
- 792KB CSS framework classes
- JavaScript bundles & chunks
- Animation wrappers & divs
- Tracking scripts & pixels
- React hydration boilerplate
- Pure semantic content
- Structured headings & data
- Accurate pricing & features
- Zero hallucination risk
- JSON-LD structured data
Three layers. One engine.
One cleaning engine. One accuracy scoring model. One edge runtime. Read, Serve, and Act all share the same protocol-level work — and that's the moat. Anyone can ship a clean Markdown API. Only Onto is building the full compatibility layer over the existing web stack, end to end, while keeping the human experience untouched.
Read
REST API + MCP server
Any AI developer turns dirty web HTML into clean, agent-ready Markdown with sub-100ms cache hits. A Firecrawl alternative with accuracy scoring, hallucination risk flagging, and MCP server support for direct integration with Claude Code, Cursor, and any MCP-compatible client.
# Firecrawl alternative curl https://api.buildonto.dev/v1/read \ -H "Authorization: Bearer $TOKEN" \ -d '{"url":"https://stripe.com"}'
Serve
Next.js SDK + edge middleware
Site owners install one line in middleware. AI crawlers get clean Markdown. Humans get the full experience, unchanged. Our own AIO score went 50 → 80. Gemini stopped hallucinating our scoring formula. Claude reads us in Markdown and explicitly confirms the MIME type.
# One line install npm install @ontosdk/next # That's it.
Act
Browser interaction — the destination of the agent web.
When agents can read, serve, and act through one engine, the agent web becomes fully programmable. Semantic intent over brittle CSS selectors — no breakage when the UI changes. Site owners opt in. The layer no one else can ship without Read and Serve first.
# Coming Q3 2026 await onto.act({ intent: "buy pro plan", context: "checkout flow" })
Same engine. Same accuracy scoring. Same edge runtime. Same auth.
Build for agents once and never re-do it.
Already speak MCP? You're already set up.
Onto ships an official Model Context Protocol server. Drop it into Claude Code, Cursor, Cline, Zed, Continue — any MCP-compatible client. Three tools, one config block, one restart.
{
"mcpServers": {
"onto": {
"command": "npx",
"args": ["-y", "@ontosdk/mcp"],
"env": {
"ONTO_API_KEY": "onto_sk_live_…"
}
}
}
}- Claude Code
- Claude Desktop
- Cursor
- Cline
- Zed
- Continue
read_url— clean Markdown for any URLscore_url— AIO score & hallucination riskread_and_score— both in one call
Why not just use Firecrawl?
Firecrawl is excellent at Layer 1. But it stops there. Onto is the only company building Read, Serve, and Act with one engine — and the only one with accuracy scoring and hallucination risk flagging at every layer.
| Capability |  Onto |  Firecrawl |  Browserbase | Vercel CF |
|---|---|---|---|---|
| HTML → Markdown (Layer 1) | Partial | |||
| Hallucination risk scoring | ||||
| AIO accuracy score | ||||
| Site-side SDK (Layer 2) | Partial | |||
| Works across all hosting | Own customers only | |||
| Browser interaction (Layer 3) | Coming | |||
| MCP server | ||||
| Control plane / analytics | Coming | Partial | Vercel only |
The infrastructure need is already proven.
Vercel, Cloudflare, Anthropic, and Y Combinator have all independently validated the thesis. Onto is how you implement it today — across any stack, any framework, any AI agent.
~99% payload reduction observed.
Serving Markdown to AI agents reduces payload size by ~99%. Complex React pages drop to a few KBs of pure content.
Read blog postCloudflare<15ms native HTML→MD at the edge.
Markdown for Agents converts HTML at the edge — positioning AI crawlers as first-class citizens of the web.
Read blog postAnthropicDirect context structure → extraction accuracy.
How context is structured — instructions, document layout, the position of the question — materially changes Claude's accuracy and recall on long inputs.
Read the post
Y CombinatorOpen RFS infrastructure for the agent economy.
YC's Request for Startups explicitly calls for developer infrastructure that makes the agent economy reliable.
Read the RFSWhat do AI agents see on your site?
Drop any URL. Get your full AI-readiness report in seconds. React Tax, semantic richness, hallucination risk, content negotiation status. Free. Instant. No signup.
Your AI representation is broken right now. We fix it.
Join the waitlist. Get the API key, the SDK, and the AIO Control Plane the day they ship. Founding pricing locked forever for early access customers.
