inetGeek MCP server

A Model Context Protocol server over inetGeek's infrastructure data: every provider and comparison on the site, and every figure carrying the URL it was read from, the sentence it was read from, and the date it was last checked against that page.

No account, no API key, and nothing you ask it is stored. Every tool is read-only except send_feedback, which records a report for the maintainer and can change nothing.

https://inetgeek-mcp.palashbagchi.workers.dev/mcp

Why point a model at this

A model asked "what does Neon's free tier include" answers from training data that may predate this week's repricing, and cannot tell you where the number came from. Every value here is read from the provider's own documentation and ships with four things: the value, the verbatim sentence it was read from, the source URL, and the date. If a figure is wrong, the excerpt makes it easy to demonstrate.

What the server will not do is more of the point than what it will. It never infers a value from silence, never fills a gap with a plausible number, and never returns a ranking without the interval around it. Where two providers cannot be told apart, it says so.

Related MCP server: whois

Connect it

Claude Code

claude mcp add --transport http inetgeek https://inetgeek-mcp.palashbagchi.workers.dev/mcp

Claude Desktop / any MCP client — add to your client's config:

{
  "mcpServers": {
    "inetgeek": {
      "type": "http",
      "url": "https://inetgeek-mcp.palashbagchi.workers.dev/mcp"
    }
  }
}

A machine-readable server card lives at /.well-known/mcp.json, and every underlying feed is indexed at /.well-known/api-catalog.

Tools

Three things worth reading the output carefully for

A rank is usually an ordering, not a finding. iScore carries an 80% interval, and a rank is only established when the posterior puts a provider above the next one at 90% or better. That is true for 4 of 222 rows. Everywhere else the ordering is real but the separation is not, and the tools say so. Report the interval with the number.

A blank is not a "no". Three different reasons a cell is empty are kept apart, because collapsing them would invent information: not documented (nobody has looked), not offered (the product has no such thing), and not stated on the vendor's page when read on a date — the last being a sourced observation, exposed through get_ledger. Only the third one is evidence, and even then it is evidence about a page, not about a product.

The lean is inetGeek's reading, not a provider's claim. compare_providers returns it with the basis it was computed from, by six published rules. It is an opinion with its working shown; pass it on that way.

Running your own

npm install
npm run dev        # wrangler dev
npm run typecheck
npm run deploy     # wrangler deploy

The server is stateless: it reads inetGeek's public JSON feeds and Markdown mirrors at request time and holds no database. Deploying your own copy gives you the same data from the same origin — useful if you want a different tool surface, different descriptions, or your own Cloudflare account in the path.

Three optional bindings. Without any of them the server still runs.

USAGE, an Analytics Engine dataset, counts calls per tool. It records the method, the tool name and the client's self-declared name — no IP address, no argument values, nothing identifying a person. Drop the analytics_engine_datasets block from wrangler.jsonc to run without it.

FEEDBACK (D1) and FEEDBACK_RATE_LIMIT back send_feedback. The row holds what the agent submitted, the time and the client's User-Agent. The IP is used only as the rate-limit key (5 a minute) and is never written. A daily cron deletes closed reports after 90 days and every report after a year. To collect feedback on your own copy:

wrangler d1 create <your-db-name>        # put the id in wrangler.jsonc
wrangler d1 migrations apply <your-db-name> --remote

Without the binding, send_feedback still registers and tells the caller that feedback is not being collected. Treat every stored message as untrusted input: it was written by an arbitrary caller of a public tool.

Data, licence and corrections

The code here is MIT. The data it serves is not part of this licence — it lives at inetgeek.com and is published for reading, with each figure attributed to the vendor page it came from.

Found a figure that is wrong or has moved, or a page of yours we recorded as silent when it is not? Send the URL: https://inetgeek.com/contact/. Corrections replace the record rather than argue with it, and the change is published at /ledger/.

• Method: https://inetgeek.com/methodology/

• How the iScore is computed: https://inetgeek.com/iscore/

• The ledger: https://inetgeek.com/ledger/