What it took to get a learning app into Claude's connector directory

Braineos has had a Claude connector for a while: a remote MCP server that lets someone’s own Claude build flashcard decks, quiz them on what’s due and record the results in their review schedule. Until this month you could only add it by pasting its address into Claude’s settings. Now it’s listed in Claude’s connector directory, so anyone can find it and connect it in a click.

Getting there was less about new features than about being explicit. Here’s what that meant.

Every tool says what kind of tool it is

The directory’s review asks each tool to declare whether it only reads or whether it can change or delete things. Our tools split cleanly: searching decks, reading a deck, listing what’s due and reading your stats only read; creating decks, adding or editing cards, attaching an image and recording a quiz all write, and deleting a card is marked as destructive. Each tool also got a short title and a description that says what it does, not how to use it.

We didn’t want those labels to drift as tools were added, so a test fails if any tool is missing its title or its read or write label. When we later added a tool that lets Claude read your learning history (accuracy by week, the cards you miss most), it went through the same check on its first commit.

Instructions that state facts

An MCP server can send its client a block of instructions. The review prefers plain facts to directions about how to behave, so ours says what the server holds and the limits that apply: how many cards a deck can have, how many decks a day, which decks are public. Claude decides how to use that; we don’t tell it how to behave.

The same rules as the website

The connector isn’t a side door. Every tool goes through the same procedures the website uses, with the same checks: a private deck stays private, a locked deck stays locked, and you can only change decks you own. On our sites where learners don’t make their own decks, the deck-editing tools simply aren’t offered, and the connector’s help page says so. We found that page describing tools a site didn’t have, and fixed it to read from the same settings as the tools themselves.

Paperwork that matters

Three things outside the code turned out to be part of the review:

• A public help page at braineos.com/claude: how to connect on claude.ai and in Claude Code, what each tool does, the limits, and how to disconnect.
• The privacy policy gained a section on the connector and names every service that receives data.
• A reviewer account with real content in it: decks, cards due for review, a few study sessions. We wrote a script of prompts that exercises every tool, ran it a day before submitting so the account looked lived-in, and gave reviewers the same script.

Sign-in uses standard OAuth, and the server publishes where to sign in, so Claude handles the whole flow.

Listed

The connector went in as a Community listing, the directory’s default. It’s at claude.ai/directory/braineos, and the Braineos announcement covers what learners can do with it.

The lesson for anyone building one: write down what each tool does and doesn’t do, enforce it with a test, and route everything through the code paths you already trust. The review is mostly a check that you’ve done that.