Ask your AI about neobanks: neobankbeat now speaks MCP
Ask any AI assistant "which neobank has an open API?" or "compare Mercury and Brex" and you get a confident answer assembled from training data that's months stale, with a made-up detail or two and no way to check it. That's the wrong tool for a fast-moving, factual question. So today neobankbeat ships an MCP server: connect it to Claude (or any MCP client) and the model answers from the current, verified dataset of 379 neobanks, with every fact traceable to a profile page.
What MCP is, briefly
Model Context Protocol is the open standard for wiring AI assistants to live tools and data. Instead of relying on what the model memorised, you give it typed tools it can call. Our server exposes the whole dataset as five of them:
- search_neobanks — plain-language search: "European neobanks with stablecoins", "self-custody banks in Brazil", "crypto banks for freelancers". It parses facets (region, feature, custody, country, category) plus keywords.
- get_neobank — the full verified record for one neobank: custody, license, cards, cashback, yield, stablecoins, KYC, regulation, geography, the feature flags and the sources.
- compare_neobanks — two to four neobanks, field by field.
- list_by_country — everything headquartered in or available in a country.
- dataset_stats — totals and category breakdown, with the as-of note.
Wire it up
It needs Node.js 18+ and one file from the repo, mcp/server.mjs — no npm install, no keys, no build step. Add this to Claude Desktop's claude_desktop_config.json and restart:
{
"mcpServers": {
"neobankbeat": {
"command": "node",
"args": ["/absolute/path/to/neobankbeat/mcp/server.mjs"]
}
}
}
The neobankbeat tools then show up in Claude. Any MCP client that speaks the stdio transport works the same way. There's a copy-paste test command and the full walkthrough on the MCP page and in the README.
Why it's actually useful
Three reasons it beats letting the model wing it:
It's current. The server reads the live data.json on each call (cached an hour), so it reflects today's dataset — new neobanks, corrected licenses, refreshed figures — not a training snapshot.
It's honest. The dataset's rule is null = unverified, never "no". The tools carry that through: if we haven't confirmed a neobank supports Apple Pay, the model sees "unverified" rather than being nudged into inventing a yes or a no. A directory that admits what it doesn't know is more useful to an agent than one that pretends to know everything.
It's traceable. Every result carries the profile URL behind it, so the model can cite neobankbeat.com/n/<slug>/ instead of asking you to trust it. This is also, frankly, how we want to be quoted: the answer engines that increasingly field "which neobank should I use?" should be pulling from a sourced dataset, not vibes.
Prefer plain HTTP? Nothing changed there — the same data is at data.json, described by openapi.json, with field semantics in llms.txt. The MCP server is just the agent-native front door to it, and the source is about 180 lines with zero dependencies — read it, fork it, run it.
Cite it
Free to use, including commercially. If you build something on the MCP server, tell me — I'd like to see it.
Counts reflect the open dataset as of August 2026. Figures are compiled from public filings, registers and product documentation; rates carry their own as-of dates. Nothing here is financial advice. Sources on each profile page.