catalog-mcp

An MCP server that turns any JSON catalog into query tools for AI agents.

Point it at a catalog URL or file — an inventory feed, a product list, the catalog.json that feedmerge publishes — and any MCP client (Claude Desktop, Claude Code, anything speaking the protocol) gets structured filtering, grouping, ranking, and schema discovery over your records.

Node 18+. Two runtime dependencies: the MCP SDK and zod.

Why

Agents are bad at big JSON files and good at tools. Hand an agent a 2 MB catalog and it will truncate, skim, or hallucinate records; hand it catalog_query with a filter grammar and it answers "cheapest record under $30k with these two features" correctly every time, reading only the records that match.

This repo is the generalized version of an MCP server I run in production: a sales-floor AI assistant queries a live inventory catalog through exactly these tools (same filter semantics, same null-price rule, same TTL cache) hundreds of times a day. The pipeline it belongs to:

vendor feed  ->  feedmerge  ->  catalog.json  ->  catalog-mcp  ->  any agent
             (guarded sync)   (versioned)      (query tools)

I run this against my own public inventory feed; the example below uses a neutral catalog so the repo stands alone.

Quickstart

git clone https://github.com/stevyf93II/catalog-mcp.git
cd catalog-mcp
npm install
npm test                                          # engine, loader, and stdio end-to-end tests

# serve the example catalog
node src/server.js --file examples/telescopes.json --key sku

Wire it into Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "my-catalog": {
      "command": "node",
      "args": ["/path/to/catalog-mcp/src/server.js"],
      "env": {
        "CATALOG_URL": "https://example.com/catalog.json",
        "CATALOG_KEY": "sku"
      }
    }
  }
}

Then ask the agent things like "what types are in the catalog and what does each cost at the low end?" and watch it compose catalog_schema, catalog_count_by, and catalog_top on its own.

Tools

All tools are read-only and idempotent, and say so in their MCP annotations.

The filter grammar

One small spec, used by query, count_by, and top:

{
  "eq":       { "type": "reflector", "goto": true },
  "min":      { "aperture_mm": 150 },
  "max":      { "price": 1000 },
  "has":      { "features": ["Parabolic Mirror", "Cooling Fan"] },
  "contains": { "name": "dobsonian" }
}

• eq — strict equality on any value, including booleans and null.

• min / max — numeric bounds. A record without a real number in a bounded field is excluded. This rule is load-bearing: in the production catalog a missing price means "call for price", and "show me units under $30k" must never surface a unit whose price is unknown.

• has — array membership; every listed value must be present.

• contains — case-insensitive substring on a string field; field "*" searches every string field in the record.

Conditions AND together. An unknown top-level key is an error that names the valid keys, because a silently ignored filter is how an agent confidently reports wrong answers.

Sorting pushes records that lack the sort field to the end, in both directions — "sort by price" shows priced records first, not a wall of nulls.

Configuration

When CATALOG_RECORDS_PATH is not set, the loader uses the document root if it is an array, or the single top-level array of objects if there is exactly one ({ "meta": ..., "items": [...] } just works). If the document is ambiguous it refuses and names the candidate keys.

On a failed refresh the server serves the last good data instead of erroring — an agent mid-task is better off with five-minute-old records than an exception — and catalog_stats reports the cache age so staleness is never hidden.

Non-goals

• Not a database. The catalog is read-only and lives in memory; if your data does not fit comfortably in a JSON file, you want a real store.

• No writes. Nothing here mutates the catalog — that is the sync pipeline's job (see feedmerge).

• No query language. Five filter keys cover what agents actually ask; anything fancier belongs in code, not in a tool schema.

License

MIT