MCP Services

Two read-only MCP (Model Context Protocol) servers, each exposing one curated internal domain wiki as tools an AI client can query directly. The underlying reference libraries are linked below.

Read-only means exactly that: nothing reachable through these connectors can add to, change, or delete either wiki. Access is gated by a key, issued per endpoint — a key for one does not work on the other. Without one, these endpoints refuse the request.

Endpoints

Tools — Avalanche Domain Wiki

• wiki_get_hot_cache Get the wiki's rolling session recap (wiki/hot.md, ~500 words): the most recent curator-agent activity and open threads. The cheapest way to check whether a question is already covered by recent work.
• wiki_get_overview Get the wiki's executive summary and source table (wiki/overview.md): page counts, source classifications (Foundational/Supplementary/Archival), and geographic coverage. Good second stop after wiki_get_hot_cache for orienting to the wiki's scope.
• wiki_get_index Get one of the wiki's catalog indexes: 'master' (every page, one-line descriptions), 'concepts' (atomic synthesis pages), 'sources' (formatted bibliography of ingested primary sources), or 'domains' (top-level and sub-domain hub pages). Use wiki_search instead if you already have a specific keyword rather than wanting to browse a category.
• wiki_search Keyword search across the wiki's markdown pages (concepts, sources, domains, entities, comparisons, questions). Returns matching page paths with short text snippets. Does NOT search library/ PDFs — only the synthesized wiki pages. Pass a returned path to wiki_get_page for the full page.
• wiki_get_page Get the full markdown content of one wiki page by its repo-relative path (e.g. 'wiki/concepts/alpha-beta-runout-model.md'). Paths aren't guessable from a topic name alone — get an exact path from wiki_search or wiki_get_index first.

Tools — UAV Survey Wiki

• uav_get_hot_cache Get the wiki's rolling session recap (wiki/hot.md): the most recent curator-agent activity, active work and open threads. The cheapest way to check whether a question is already covered by recent work.
• uav_get_overview Get the wiki's executive summary (wiki/overview.md): what the wiki covers, what is out of scope, and how its folders are organized. Good second stop after uav_get_hot_cache for orienting to the wiki's scope.
• uav_get_index Get one of the wiki's catalog indexes: 'master' (every page, with tags and source citations), 'concepts' (technical procedures, parameters and methods), 'entities' (software, hardware, sensors and tools), 'domains' (end-to-end workflow overviews), 'sources' (bibliography of ingested documents, grouped by type), or 'taxonomy' (the controlled tag vocabulary — useful for picking search keywords). Use uav_search instead if you already have a specific keyword rather than wanting to browse.
• uav_search Keyword search across the wiki's markdown pages (concepts, entities, domains, sources, questions, meta). Returns matching page paths with short text snippets. Does NOT search library/ source documents — only the synthesized wiki pages. Pass a returned path to uav_get_page for the full page.
• uav_get_page Get the full markdown content of one wiki page by its repo-relative path (e.g. 'wiki/concepts/gnss-preprocessing-ppk.md'). Paths aren't guessable from a topic name alone — get an exact path from uav_search or uav_get_index first.

Reference libraries

Each wiki is built from a library of source documents held in its private repository. Browse or download from the folder directly — the listing is live, so documents added later appear without this page changing.

Both repositories are private — you'll need to be signed in to GitHub with access to the Alpine Solutions organisation. Avalanche papers are stored with Git LFS; GitHub's web interface will download them normally, and a git clone needs a Git LFS client to fetch the file contents.

Documents in these libraries are the property of their respective publishers and are held for internal reference. Redistribution outside Alpine Solutions isn't covered by that.

Querying the wikis

Once a connector is attached, just ask in plain language — the model picks the tools. Naming the wiki in your question helps when both are connected.

"Check the avalanche wiki: what does it say about alpha-beta runout models?"
"Search the UAV wiki for RINEX header verification and quote the relevant page."
"What's changed in the avalanche wiki recently?"          → reads the session recap
"List what's in the UAV wiki."                            → reads the master index

Getting good answers

• Work from recap → index → search → page. Ask for the recap or an index first when you don't know what's there; go straight to search when you have a keyword.
• Ask for the source. Every page cites where its content came from, so "and cite the page you used" gets you a checkable answer.
• Searches match keywords, not meaning. If "snow depth accuracy" finds nothing, try the terms the wiki itself would use — the UAV wiki's taxonomy index lists its controlled vocabulary.
• Neither server searches the reference libraries — only the synthesized wiki pages. For the underlying documents, use the library links above.

Querying directly over HTTP

For scripting or debugging, the endpoints speak JSON-RPC. Both headers below are required; the response comes back as an event stream.

curl -X POST ENDPOINT \
  -H "Authorization: Bearer KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Call a tool by name, passing its arguments:

curl -X POST ENDPOINT \
  -H "Authorization: Bearer KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"wiki_search","arguments":{"query":"runout","limit":5}}}'

Connecting

Throughout, ENDPOINT is the URL from the table above and KEY is your issued key.

Claude Code

One command per machine, then /mcp to confirm it connected:

claude mcp add --transport http --scope user wiki ENDPOINT \
  --header "Authorization: Bearer KEY"

Claude Desktop

Needs Node; Desktop can't send headers itself, so it uses mcp-remote. Add to claude_desktop_config.json, then quit and reopen Desktop fully:

{
  "mcpServers": {
    "wiki": {
      "command": "npx",
      "args": ["mcp-remote", "ENDPOINT", "--header", "Authorization: Bearer KEY"]
    }
  }
}

claude.ai and mobile

Settings → Connectors → Add custom connector. Set the URL to ENDPOINT, leave the OAuth fields empty, and add a request header Authorization with value Bearer KEY.

ChatGPT

Not currently supported. ChatGPT custom connectors authenticate only with OAuth or no authentication. They cannot send an Authorization header with a fixed key, so both endpoints refuse ChatGPT's requests.

To use the wikis from OpenAI tools, use one of these instead:

• OpenAI API: attach the server as an mcp tool (next section).
• Codex: add the server to ~/.codex/config.toml and set the key in an environment variable, here WIKI_KEY. Codex sends it as the bearer token:

[mcp_servers.wiki]
url = "ENDPOINT"
bearer_token_env_var = "WIKI_KEY"

OpenAI API

In the Responses API, attach the server as an mcp tool. Note that authorization takes the bare key — no Bearer prefix, unlike every other client here. The value isn't stored by the API, so send it on each request:

{
  "type": "mcp",
  "server_label": "avalanche_wiki",
  "server_url": "ENDPOINT",
  "authorization": "KEY"
}

VS Code — GitHub Copilot

Add the server to mcp.json. Copilot's own guidance is to avoid pasting the key into the file; an input variable prompts for it instead and keeps it out of source control:

{
  "servers": {
    "wiki": {
      "type": "http",
      "url": "ENDPOINT",
      "headers": { "Authorization": "Bearer ${input:wikiKey}" }
    }
  }
}

Microsoft Copilot Studio

Add the MCP server with ENDPOINT as its URL and choose API key authentication. Set the location to Header, the header name to Authorization, and the value to Bearer KEY.

Repeat with the other endpoint to connect both wikis in one client.

Access

Treat your key like a password — not in tickets, chat, or commits. For access, rotation, or a fault on either endpoint, contact the Alpine Solutions wiki administrator.