An MCP Server for Every Hill in Britain and Ireland

The parkrun MCP server proved that zero configuration is worth chasing: no API keys, no OAuth, install and ask. This one goes a step further. There’s no API at all. The entire dataset ships inside the npm package.

It answers questions like “which Munros are within 20km of Fort William?”, “tell me about Helvellyn”, and “how many Marilyns are there?”. Every hill in Britain and Ireland, 21,576 of them, across 31 classifications.

What it does

Five tools:

• find_nearby_hills takes a latitude and longitude and returns hills sorted by distance, optionally filtered to one classification (only Munros, only Wainwrights)
• search_hills is a ranked name search that handles the Gaelic diacritics in names like Sgùrr nan Gillean
• get_hill_details gives you height, drop, grid reference, coordinates, and every list a hill belongs to
• get_list returns a whole classification sorted by height, with optional height filters
• get_hill_stats breaks the database down by classification and country

Setup is one command:

claude mcp add hills -- npx -y @michaelhutchinson/hills-mcp-server

The data source

The Database of British and Irish Hills is a volunteer-maintained dataset that’s been going for over twenty years. It records every hill on these islands with surveyed heights, prominence, grid references, and membership of every hill list you’ve heard of and plenty you haven’t: Munros, Corbetts, Wainwrights, Marilyns, HuMPs, Dodds. It’s licensed CC-BY 4.0, so you can build on it as long as you credit it. Every response from the server carries the attribution.

The database updates a few times a year. That release cadence is what makes bundling work: a build script processes the official CSV into a trimmed JSON that gets committed and shipped in the package. Lookups are instant, nothing touches the network, and the server works on a train in the Highlands, which is thematically appropriate. When a new DoBIH version lands, I rerun the script and publish a patch release.

Two gotchas in the data

Two things bit me that are worth knowing if you ever work with this dataset.

First, the CSV has a Classification column listing each hill’s codes, and I assumed it was complete. It isn’t. Four lists (Murdos, Trail 100, Carns, Binnions) only exist as separate 0/1 flag columns. My stats tool confidently reported zero Murdos in Scotland until I actually read its output while testing. There are 441.

Second, twelve hills carry codes with an equals suffix, like Ma=, marking a summit of equal height to its parent. Filter on the exact string Ma and the Marilyns among them silently vanish. The right count is 2008, not 2007. Nobody would ever notice that bug from the output. It only surfaced because a code review queried the raw data and checked the counts.

Error messages are an API for models

One bug taught me something MCP-specific. When a classification didn’t resolve, the server replied with the list of valid names. Two of those names didn’t themselves resolve, because my alias table was hand-written and incomplete. So the error message told the model to retry with a value the server would reject, which is an infinite loop with extra steps.

The fix was to derive the aliases from the classification data itself, so the suggestion list and the resolver can’t drift apart. The general rule: in an MCP server, an error message isn’t documentation, it’s an instruction the model will follow. If your error suggests a value, your code must accept that value.

Try it

claude mcp add hills -- npx -y @michaelhutchinson/hills-mcp-server

Source on GitHub, package on npm. Hill data from the Database of British and Irish Hills, CC-BY 4.0.