RipSeek, one of our products, has a remote MCP server: AI apps such as Claude, ChatGPT and Cursor can ask it about Collector Crypt card packs (odds, the most valuable cards in each pack, recent pulls) and get live answers. Within a day of publishing, it was in the official MCP Registry and on Glama, where it scored A (3.9 out of 5) on tool quality, with a healthy status. Here is what mattered.
The server: small and boring on purpose
- Remote, Streamable HTTP, stateless. One
POST /mcpendpoint on a Cloudflare Worker that answers JSON-RPC. No sessions to store, so it scales like any other Worker route. - Read-only, no sign-in. Every tool only reads. Nothing can be bought or sold through it; to act, the user goes to the product and approves in their own wallet. Read-only tools are also the ones agents and directories trust most.
- Six tools, each with one job.
list_packs,get_pack,search_cards,get_card,list_recent_pulls,estimate_budget.
What a good tool description looks like
Glama scores each tool's definition, not your marketing. Its criteria are published as the Tool Definition Quality Score (TDQS); the top grade is A, from 3.5 up. The descriptions that score well, and that agents actually pick correctly, follow one pattern:
- First sentence: what it does and what comes back.
- When to use it, and when to use another tool instead. Agents choose between your tools by reading these; say it plainly.
- What it touches and how fresh the data is. "Read-only; pools are rescanned every few minutes."
- An example in every parameter. Not just a type:
"Words that must all appear in the card name, any order, e.g. 'umbreon vmax' or 'psa 10 rookie'." - An output schema, so clients that validate structured results know what they'll get. We keep ours loose (extra fields allowed) so a stricter client never rejects a reply.
- Annotations:
readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: true. - Verb-first names:
list_recent_pulls, notrecent_pulls. When we renamed tools, the old names kept working as aliases so no client broke.
A real one, get_card:
Shows one Rare or Epic card that is in a pack pool right now: name, insured value, picture, and every pack it can come from with that pack's price and the chance per pack of pulling this exact card. Use it after search_cards, passing the nft field of a result; use search_cards when you only have a card name. Read-only; fails with a clear message if the card has just been pulled.
Glama's first review also flagged gaps in coverage, not just wording. It suggested making one tool's output clearer and adding a way to look up a single card; we added get_card and filters to list_recent_pulls. Then we rewrote all six descriptions against the pattern above.
Errors are part of the interface
When a tool can't answer, return a normal result with isError: true and a sentence the model can act on (for example "There's no pack with the code pokemon_5. Call list_packs to see the codes."), not a protocol error. The agent reads it, fixes its call and tries again, instead of giving up.
Getting listed, in the right order
- The official MCP Registry first. Publish a
server.json(name, description, remote URL, icon) with Anthropic'smcp-publishertool. The name is a namespace you prove you own:io.github.<org>/<server>through GitHub, orcom.yourcompany/<server>through a DNS TXT record on your domain. Use your own domain if you have one: it tells everyone the server is official. - Glama and other directories pick it up from the registry. Ours appeared on Glama within a day, with the score and a health check. Claim it on Glama to see its analytics.
- Then the awesome lists. Some require the Glama badge, so they come last. Read each list's contributing rules: one wanted descriptions under 120 characters, another a specific place in a data file.
- The Claude connector directory is submitted by the company itself, from its Claude admin settings, and needs OAuth if the server needs sign-in, plus a privacy policy and read-only or destructive labels on every tool. It doesn't accept crypto software in general.
What we'd tell an API company
If you already have a public API, an MCP server is mostly a translation job: pick the 5 to 10 things people actually ask your product, write each as a tool with the pattern above, keep it read-only to start, and publish it under your own domain before someone else wraps your API under theirs. There are already hundreds of unofficial servers in the registry built on other companies' APIs.
We build these for other teams: on your existing API, on your own domain, listed in the registry and the main directories, from US$500. See our services page.