Camptocamp MCP Server

mcp-camptocamp

An MCP server that lets an LLM look up mountain routes, summits, huts and trip reports on Camptocamp.org instead of guessing them.

What it does

Ask an LLM for the altitude of a summit or the rating of a ski tour and it may answer from memory, and get it wrong. This server gives it read-only access to Camptocamp.org, the collaborative mountaineering site: routes with their ratings and descriptions, summits and huts with their altitudes, trip reports with the conditions people found, areas, guidebooks and articles. Every tool output comes from the public Camptocamp API, with the camptocamp.org link to cite, and a value Camptocamp does not have is left out, never guessed. It needs no account and no API key.

An example conversation from October 2026, run against the published server (v1.3.0); the answer's table is cut to its first 5 rows:

Prompt: Give me the top 10 of most climbed routes on l'aiguille verte. With a very concise response

Called camptocamp 20 times

Top 10 routes on the Aiguille Verte, ranked by the number of camptocamp outings logged since the site began:

#	Route	Grade	Outings
1	Couloir Whymper	AD+	241
2	Couloir Couturier	D	92
3	Arête du Jardin	D-	33*
4	Arête des Grands Montets	D	24
5	Arête du Moine	AD+	22*
…	5 more rows

* Two counts are inflated. The Arête du Jardin figure includes traverses that stop at the Aiguille du Jardin without reaching the Verte. The Arête du Moine figure includes outings where it was only the descent route.

These are camptocamp reports, not actual ascents.

Quick start

The server runs on your machine, started by your MCP client over stdio. Run it with npx, which needs Node.js 22 or later, or with Docker.

npx. Clients configured with an mcpServers JSON object, such as Claude Desktop (claude_desktop_config.json), Gemini CLI (settings.json) or Cursor (.cursor/mcp.json), take this entry:

{
  "mcpServers": {
    "camptocamp": {
      "command": "npx",
      "args": ["-y", "@olaurendeau/mcp-camptocamp"]
    }
  }
}

In Claude Code, one command adds it for every project:

claude mcp add --transport stdio --scope user camptocamp -- npx -y @olaurendeau/mcp-camptocamp

Docker. Use this entry instead; the image runs on amd64 and arm64:

{
  "mcpServers": {
    "camptocamp": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "ghcr.io/olaurendeau/mcp-camptocamp:latest"]
    }
  }
}

The first start downloads the package or the image, which can outlast a client's startup timeout. Getting started shows how to download it beforehand, a smoke test, and the Docker command for Claude Code; Troubleshooting helps when the server does not show up.

The server is listed in the official MCP registry as io.github.olaurendeau/mcp-camptocamp.

Supported clients

The support matrix lists the clients covered by the client pages and the agent SDK guide, with the date each page was last checked against the vendor's docs.

Tools

15 read-only tools: one search and one detail tool per kind of Camptocamp document, a shortcut for a user's outings, get_outings, to read several outings in one call, and outing_stats, to count outings by month, year or condition (these two from v1.4.0). Search results give IDs; the get_* tools take an ID. Every tool takes lang, the language of titles and texts (fr by default; v1.3.0 or later).

Each tool page gives its inputs, generated from the registered schema, its output format, a real example and its limits: see the tool reference.

Guides

• Using the tools with an LLM: which tools to chain for a region, a summit altitude, a hut, recent conditions or guidebooks; what each output line means; and a real June ski-tour example.
• System prompt: a prompt to paste into your agent, so the model quotes Camptocamp, cites it and says when a value is missing.
• Self-hosting over HTTP: host an instance behind a secret token for clients that only take a URL, such as Claude.ai and Vibe Work (v1.4.0 or later).
• Documentation index: every page, by topic.

Development

Everything runs in Docker, through the Makefile: no local Node.js is needed. make check runs the same steps as the CI checks job: format, lint, type check, tests with enforced coverage thresholds (95 % of lines, functions and statements, 90 % of branches), build and hook tests. Every change lands through a reviewed pull request: see CONTRIBUTING.md (in French), and Development for the make targets, the live API contract tests, releases and the stack.

License

MIT