# Onto — AI is reading your website wrong. > Onto serves clean Markdown to AI agents instead of HTML noise. Same content, right format. One line of code. No changes to your site. **Source:** /mcp-server **Extracted:** 2026-05-20T20:59:17.424Z --- LiveProduct · MCP Server ## Onto, in any _MCP client._ The official Model Context Protocol server for Onto. Drop it into Claude Desktop, Cursor, Cline, Zed, Continue or any other MCP client. Three tools — `read`, `score`, `search`. One config block. One restart. Done. [Get an API key](https://app.buildonto.dev/signup)[MCP install guide](https://docs.buildonto.dev/mcp) 3 Tools exposed 8+ MCP clients supported 1 config Install footprint 01 // Install ### One config block. One restart. Done. The Onto MCP server runs through `npx` — no global install, no infrastructure to maintain. Paste the block, restart your client, the tools appear. claude\_desktop\_config.json { "mcpServers": { "onto": { "command": "npx", "args": \["-y", "@ontosdk/mcp"\], "env": { "ONTO\_API\_KEY": "onto\_sk\_live\_…" } } } } \# Restart Claude Desktop. That's the whole install. 01**Drop in** a single `mcpServers` block. `npx` handles install — no global setup, no infrastructure. 02**Restart** your client. The Onto tools appear in the tool list with their names, descriptions, and JSON schemas auto-populated. 03**Use them** from chat. Every call is metered against your Onto API key, no separate billing or quota. 02 // Tools exposed ### Three tools your agent already knows what to do with. Each tool ships with its JSON schema and docstring auto-populated from the protocol. Your client sees them, your model invokes them, you get billed per call. Tool01 onto.read (url: string) → { markdown, score, insights, hash } Fetch any public URL, return clean Markdown with the AIO accuracy score and hallucination-risk flags inline. Tool02 onto.score (url: string) → { score, insights, penalties, recommendations } Score a URL without retrieving its content — useful for deciding whether a page is worth feeding to your model at all. Tool03 onto.search (query: string) → { results: { url, score, snippet }\[\] } Search the agent web for URLs matching a query, ranked by AIO accuracy score rather than raw popularity. 03 // Compatibility ### Works with every MCP client we've tested. If a client implements the Model Context Protocol, it can talk to Onto. We test the official Anthropic clients and the major third-party IDEs on every release. Client01 Claude Desktop Official Anthropic client Client02 Claude Code Anthropic's CLI Client03 Cursor Cursor MCP Client04 Cline Cline (VS Code) Client05 Zed Zed's MCP runtime Client06 Continue Continue.dev Client07 Windsurf Codeium's IDE Client08 Goose Block / Square's agent 04 // Lifecycle ### From config to tool call. Step 01 Add to client config Paste one mcpServers block with your API key into Claude Desktop / Cursor / Cline / Zed config. claude\_desktop\_config.json Step 02 Restart client Client picks up the new server. Tools self-describe — no manual schema entry. one restart Step 03 Use the tools Reference them by name in chat or let the model invoke them through native tool-use. onto.read · onto.score · onto.search Step 04 Metered per key Every tool call hits the same /v1 endpoints as the REST API and counts against your Read tier. shared quota 05 // Compared ### Onto MCP vs rolling your own. You can absolutely build your own MCP server. Most teams that do end up rebuilding the same surface area: auth, tool schemas, rate limits, deploys. We did all of it once. Onto MCP Roll-your-own Install ✓One mcpServers block + npx Spin up server + tunnel + auth Tools shipped ✓read · score · search Whatever you write Hosting ✓Managed · zero infra You host & monitor Authentication ✓API key in env Roll your own Schema discovery ✓Auto-populated Manual JSON Schemas Updates ✓npm — no redeploy Ship a new build Cost to operate ✓Free below 1,000 calls/mo Whatever your VPS costs 06 // Pricing ### MCP calls share your Read quota. Tool calls go through the same /v1 endpoints as the REST API, so one subscription covers both surfaces. Free forever for 1,000 calls / month. Free $0forever 1,000 calls/mo Counts against shared Read quota Starter $9/ month 10,000 calls/mo Credit packs cover overage GrowthMost picked $49/ month 100,000 calls/mo Priority queue · 20 concurrent Scale $250/ month 500,000 calls/mo Shared Slack · 50 concurrent MCP and REST API share a single quota — pay once, use either surface freely. [Full pricing breakdown](/pricing) 07 // Common questions ### Before you paste the config. 01Do I need an Onto account to use the MCP server? Yes — you need an API key to authenticate. The Free tier covers 1,000 calls per month forever and never requires a card. 02Which MCP clients are supported? Anything that speaks the Model Context Protocol: Claude Desktop, Claude Code, Cursor, Cline, Zed, Continue, Windsurf, Goose, Roo Code. If a client supports MCP, it supports Onto. 03Do MCP calls count against my Read API tier? Yes. Tool calls hit the same /v1/read and /v1/score endpoints as the REST API, so a single quota covers both surfaces. No duplicate billing. 04Is the MCP server open source? The npm package `@ontosdk/mcp` is source-available — you can read the wire format and how each tool is constructed. The cleaning engine and scorer stay managed. 05How do I rotate my API key? Rotate in [the dashboard](https://app.buildonto.dev) — update the `ONTO_API_KEY` env value in your client config and restart. We don't revoke old keys automatically, so you can dual-run during the swap. 06What happens if a tool call exceeds my quota? The call returns a structured error your client renders inline — no silent failure, no surprise charges. Top up with a credit pack or upgrade your tier to keep going. Ready to plug in ### Three tools. Any MCP client. Get an API key, paste one config block, restart your client. Free tier covers 1,000 calls / month forever. [Get an API key](https://app.buildonto.dev/signup)// Works in Claude, Cursor, Cline, Zed…