# Rubricary

> A versioned registry of system prompts with an automatic eval battery and a
> runtime delivery API. Every prompt is a git repository with versions, eval
> cases and a score against several models.

Rubricary exists so a team stops keeping its production prompt inside the code,
in an environment variable, or in a spreadsheet, and starts treating it as what
it is: a versioned, tested, reviewable artifact.

## For agents

Rubricary runs an MCP server, so an agent can search and use the catalog without
its user leaving their editor.

- MCP endpoint (Streamable HTTP): /api/mcp
- Tools: `search_prompts`, `get_prompt`, `get_evals`, `validate_prompt`
- No key required for the public catalog. `validate_prompt` shares the same rate
  limit as the website.

Configuration to copy:

```json
{
  "mcpServers": {
    "rubricary": {
      "type": "http",
      "url": "https://app.rubricary.com/api/mcp"
    }
  }
}
```

## Main resources

- /  — signed-in home: shortcuts to every surface, plus top-rated, most
       starred and recently added prompts
- /prompt — the catalog, searchable by sector, use case, locale, jurisdiction,
       model and license
- /validate — score any system prompt against an explicit rubric, no account
- /prompt/{org}/{name} — one prompt: rendered text, variables, license, version
- /prompt/{org}/{name}/evals — its eval cases and their results
- /prompt/{org}/{name}/scores — score and cost per 1,000 calls, per model
- /prompt/{org}/{name}/versions — history and the diff between any two versions

## Using a prompt

```
npx rubricary add support/refund-request
```

Every public prompt is a git repository and can be cloned directly.

## Licensing

Each prompt carries its own license, shown on its page and returned by
`get_prompt`. The license covers the prompt text; it says nothing about what a
model generates from it.