Machine-readable tools, in plain terms

An assistant answers from what it learned. A tool lets it go and get something it could not know — a live price, the state of a system, a document written after training ended. The Model Context Protocol is the common way to offer those tools, so that any assistant can call any tool without a bespoke integration for each pair.

A server, in this sense, is a small program that publishes a handful of named actions and describes what each one takes and returns. Nothing about it is specific to crypto, or to any subject.

A server.json is the listing card. It names the server, says where it lives and what it offers, and it is what the official registry indexes. The registry held 9,652 server records as of 2026-05-24 ¹. Other directories largely derive from it, so one listing tends to appear in several places. Submitting requires proving you own the namespace you are listing under, which means an account.

# what does this server offer?
curl -sS -X POST "$SERVER/mcp" \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# call one of the tools it listed
curl -sS -X POST "$SERVER/mcp" \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":
       {"name":"search","arguments":{"q":"stablecoin","limit":5}}}'

What we found in the registry on 2026-08-24

We searched it, expecting a crowded shelf. A scan of 100 servers returned by a search for "crypto", filtered for anything that explains, teaches, defines or guides, returned zero ³. A search for "education" returned 15 servers, all of them about schools, universities and lesson plans, none about this subject ³. "Glossary" returned zero. "Explain" returned 25, none of them crypto ³.

Everything in the crypto category is price data, market data or on-chain actions. One safety-screening server exists for a single chain, and it also executes trades. A screening tool that can move funds is a different proposition from one that only reads.

The obvious idea here is the wrong one

An empty shelf reads as an opening. It reads equally as a shelf nobody stocked because nobody wanted the product, and in this case the second reading is the right one.

A tool that defines "liquidity pool" competes with the model that is calling it. The model already knows, answers faster, and needs no dependency. Building the dictionary adds latency and a thing to keep alive, in exchange for nothing the caller could not already do.

What a tool call genuinely earns its place for is narrow: facts from after the model's training cutoff, live state that changes by the minute, and the specifics of your own systems — how your own receipts are structured, how your own proofs verify. That last category is the only one most teams can supply better than anyone else, and it is the one worth building.

The same logic applies to llms.txt, the file that tells language models what a site holds. Across 26 crypto domains probed on 2026-08-24, 14 carried a real one, and every single carrier was developer documentation or infrastructure ⁴. Every media and explainer site returned nothing. Ship the file because it costs one file. Do not plan around it being read.