The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Onto listing page.
The official Onto Model Context Protocol server. Add clean web content reading and AI-readability scoring to Claude Code, Cursor, Cline, Zed, or any MCP-compatible AI client.
Onto's MCP server exposes these tools to any AI agent:
read_url — Read any URL and get back clean, agent-ready Markdown (typically 10× smaller than raw HTML)score_url — Get the AIO (AI-readability) score for any URL with a breakdown of what helps and what hurts AI consumptionread_and_score — Both at once: clean content plus quality assessment so the agent knows how much to trust the sourcebatch — Read, score, or extract many URLs (an explicit list or a whole site) in one callmap_site — Discover a site's URLs via its sitemap, without reading themextract_data — Return the structured data a page already declares (JSON-LD, OpenGraph, meta)Every tool response ends with a one-line ⚡ Onto report — reduction, tokens saved, and AIO score — so the value is visible on every call.
This is the official MCP wrapper for the Onto Read API. It's a thin, deterministic layer: Onto cleans and scores with a rule-based engine (no LLM in the loop), so your agent's own model never has to parse raw HTML.
When AI agents read websites today, they parse hundreds of KB of React noise to find a few KB of actual content. This burns tokens and causes hallucinations.
Onto strips the noise server-side, returns the agent-ready format, and reports a confidence score for the source. One tool call, one accurate answer.
Sign up at app.buildonto.dev and create an API key at Read → Keys.
Free tier: 1,000 credits / month. No credit card.
Add to your Claude Code MCP config:
Restart Claude Code. The Onto tools (read_url, score_url, read_and_score, batch, map_site, extract_data) will appear in the available tools list.
Add the Onto server to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (this repo). Or use the UI: Customize (sidebar) → MCPs → add a custom server, same JSON.
Fully quit Cursor (Cmd+Q / Ctrl+Q) and reopen.
See examples/ for Cline, Zed, and Continue configs.
In Claude Code or Cursor, try:
Read https://stripe.com/pricing using Onto and summarize the pricing tiers.
The agent calls read_url, gets clean Markdown, and returns an accurate summary without parsing hundreds of KB of layout HTML.
read_urlReturns clean Markdown for a URL with metadata about the extraction (sizes, reduction %, cache state).
Input:
| Field | Type | Required | Description |
|---|---|---|---|
url | string | yes | Publicly accessible HTTP(S) URL |
fresh | boolean | no | If true, bypass cache (default: false) |
score_urlReturns the AIO (AI-readability) score for a URL — 0-100 with a letter grade, hallucination risk, and a structured list of penalties / benefits / recommendations.
Input:
| Field | Type | Required | Description |
|---|---|---|---|
url | string | yes | URL to score |
read_and_scoreReturns clean Markdown plus the AIO score in one call. Recommended default for agentic workflows.
Input: same as read_url.
batchProcess many URLs in one call. N credits — one per URL in the list after the list is known. Failed URLs (ok: false) are refunded. Default mode read-and-score is still N, not 2N. Give an explicit list or a base URL whose pages are auto-discovered.
Input:
| Field | Type | Required | Description |
|---|---|---|---|
urls | string[] | one of | Explicit list of URLs (max 50). Use this or site. |
site | string | one of | Base URL whose pages are auto-discovered via sitemap. Use this or urls. |
mode | "read" | "read-and-score" | "extract" | no | What to do per URL (default "read-and-score") |
limit | number | no | Site mode only: max pages to discover (default 25, max 50) |
map_siteDiscover a site's URLs (sitemap → on-page links) without reading them. 1 credit per call (not per discovered URL) — use it to plan which pages to read or batch next.
Input:
| Field | Type | Required | Description |
|---|---|---|---|
url | string | yes | Base URL of the site to map |
limit | number | no | Max URLs to return (default 100, max 1000) |
extract_dataReturn the structured data a page already declares — JSON-LD, OpenGraph, and meta tags — plus the AIO score. Deterministic; no fields are inferred by a model.
Input:
| Field | Type | Required | Description |
|---|---|---|---|
url | string | yes | URL to extract structured data from |
| Tier | Monthly credits | Price |
|---|---|---|
| Free | 1,000 | $0 |
| Starter | 10,000 | $9 |
| Growth | 100,000 | $49 |
| Scale | 500,000 | $250 |
| Enterprise | Custom | Contact sales |
MCP tool credits match the live API meter:
| Tool | Credits |
|---|---|
read_url | 1 |
read_and_score | 1 |
score_url | 0 (free) |
extract_data | 1 |
map_site | 1 per call (not per discovered URL) |
batch | N — one per URL in the list after the list is known; failed URLs (ok: false) are refunded. Default mode read-and-score is still N, not 2N |
Connect / OAuth is not a debit.
Manage your subscription at app.buildonto.dev/read/billing. Credit packs ($5–$200) are available for overflow once you're on a paid tier.
Environment variables:
ONTO_API_KEY (required) — Your Onto API key from app.buildonto.dev/read/keysONTO_API_BASE (optional) — Override the API base URL (default: https://api.buildonto.dev)Verify your key at app.buildonto.dev/read/keys. If you recently rotated keys, your MCP config may have a stale value.
You've used your monthly allotment. Upgrade at app.buildonto.dev/read/billing or wait for the monthly reset. Paid tiers can also top up with credit packs.
npx is on your PATHThe target site may be slow or unreachable. Onto's request timeout is 15 seconds. Retry, or try a different URL.
Onto is the compatibility layer for the agent web. Three products on one engine:
Built for AI agents reading the web. Built so they read it correctly.
MIT