Rubricary runs an MCP server at /api/mcp — Streamable HTTP, stateless, one JSON-RPC request per call. Point an agent (Claude Code, Cursor, or anything else that speaks MCP) at it with:
json
1{
2"mcpServers":{
3"rubricary":{
4"type":"http",
5"url":"https://app.rubricary.com/api/mcp"
6}
7}
8}
No key is required for the public catalog. Four tools are exposed:
Tool
Does
search_prompts
Free-text search over the catalog, filterable by sector, use case and locale.
get_prompt
One prompt's rendered text, variables, license and current version.
get_evals
Its eval cases and score per model, with cost per 1,000 calls.
validate_prompt
Score an arbitrary system prompt. Goes through the same guard chain as /validate — see Security for the rate limit it shares with the website.
A raw tools/call request looks like this:
json
1{
2"jsonrpc":"2.0",
3"id":1,
4"method":"tools/call",
5"params":{
6"name":"get_prompt",
7"arguments":{"name":"support/refund-request"}
8}
9}
A plain GET /api/mcp describes the server by hand, if you'd rather poke at it with curl before wiring up a client.