Model Context Protocol
machinesloveyou measures how a site is found — by search engines and by AI answer engines — and serves those measurements to your agent over MCP. Connect it once and your terminal can ask anything about your visibility, read a full specialist review, and go fix what it finds. Read-only, scoped to your account.
claude mcp add --transport http machinesloveyou https://machinesloveyou.com/mcp \ --header "Authorization: Bearer seo_live_YOUR_TOKEN"
Any MCP client works — the command above is for Claude Code. Create the token in your account settings; it is shown once, stored as a hash, and revocable at any time. The full reference, with every tool and its parameters, is in the docs.
Which queries lost clicks last month and on which pages. Which rivals gained the positions you lost. Which of your pages Google declared indexable and then never indexed. Your agent reads the measurements and answers in your own terms — no dashboard to learn, no export to reconcile.
The agent that reads the finding is the one holding your codebase. A truncated title, a price stated differently on two pages, a canonical pointing at the wrong URL: it reads the proof, opens the file and writes the change. Nothing gets copied between a report and an editor, which is where findings usually go to die.
Not what you rank for — what a model says when someone asks about your market, which sources it read to say it, and who it named instead of you. That is measured by asking the engines and recording the answers, so your agent can reason over what actually came back rather than over a guess about how models work.
31 lenses, each a different specialist reading your site: technical SEO, AI visibility, authority, conversion, e-commerce, retention, pricing. Your agent pulls one in full — score, per-dimension breakdown, every finding with the proof behind it — and turns it into work.
An agent reading SEO data is reading the open web: page titles, search snippets, forum threads, AI answers. That content is written by strangers, and some of it will try to talk to your agent. The server is designed for that.
42 datasets of measured state — rankings and queries, crawl and indexability, speed and Core Web Vitals, backlinks and mentions, competitor movement, AI citations and the sources behind them. Datasets ending in _history carry the time series. An empty answer means not measured yet, never that the site is fine.
31 lens reviews, each a specialist reading of the site against a scored rubric. The rule they all obey: no finding without proof — a quote checked against the page text that was actually read, or a value checked against a dataset that was actually collected. Findings whose proof does not hold are discarded before the report is written, which is why a review can come back saying it does not have enough to say.