storywrangler-mcp 0.1.0

MCP server exposing Storywrangler documentation and the dataset registry to AI agents

storywrangler-mcp

MCP server that lets AI agents work with the Storywrangler platform. Modeled on the Svelte team's @sveltejs/mcp and vcsi-starter's @the-vcsi/scrolly-mcp: the server holds no content — it fetches documentation from the docs site's llms.txt exports and dataset metadata from the live registry API, so nothing can drift.

Tools

Tool	Source	Purpose
list-sections	docs site /sections.json	Discover documentation sections (guides + per-tag API reference)
get-documentation	docs site /{slug}/llms.txt	Fetch section markdown, with fuzzy matching
list-datasets	API /registry/	List registered datasets with queryable dimensions
get-dataset	API /registry/{domain}/{dataset_id}	Ground-truth metadata: level_order, filter_values, availability
validate-submission	local (storywrangler-schemas)	Dry-run a DatasetCreate payload: real Pydantic contract, mirrored server guards, conflation lints, on-disk hive layout checks when reachable

Usage

Remote (streamable HTTP) — the backend mounts this same server at /mcp (see backend/app/main.py). No install needed:

{
  "mcpServers": {
    "storywrangler": {
      "type": "http",
      "url": "https://storywrangler.uvm.edu/mcp/"
    }
  }
}

Keep the trailing slash — /mcp works but costs a 307 redirect per request. On the server, set STORYWRANGLER_DOCS_URL/STORYWRANGLER_URL to localhost addresses so tool fetches don't loop through the public proxy. Note: remote clients validate TLS strictly; this endpoint is unusable externally until the uvm.edu certificate mismatch is fixed.

Local (stdio) — works offline and pins a version:

{
  "mcpServers": {
    "storywrangler": {
      "command": "uvx",
      "args": ["storywrangler-mcp"]
    }
  }
}

CLI mode

Every tool also works as a shell command — the fallback for subagents and CI where no MCP connection exists:

uvx storywrangler-mcp list-sections
uvx storywrangler-mcp get-documentation register querying
uvx storywrangler-mcp list-datasets
uvx storywrangler-mcp get-dataset wikimedia ngrams --full
uvx storywrangler-mcp validate-submission payload.json   # '-' reads stdin

validate-submission exits non-zero on blocking errors, so a pipeline can gate its submit step on it (add --no-disk to skip layout checks).

Configuration

Env var	Default	Meaning
STORYWRANGLER_DOCS_URL	https://storywrangler.uvm.edu	Docs site base URL
STORYWRANGLER_URL	https://storywrangler.uvm.edu	API base URL (same var the SDK uses)
STORYWRANGLER_INSECURE	unset	Set to 1 to skip TLS verification (the uvm.edu deployment currently serves a mismatched certificate — fix pending)

Point both at http://localhost:5173 / http://localhost:8000 for local development.