MCP & agents
Tool reference
The server exposes four tools. Each one calls the same code as the website, so an agent sees the same results a person would. Every argument is a string.
search_prompts
Searches the public catalog.
| Argument | Required | Description |
|---|---|---|
| query | no | Free text matched against name, description and sector. |
| sector | no | e.g. Support, Sales, Engineering, Legal. |
| use_case | no | e.g. Drafting, Classification, Extraction, Summarization. |
| locale | no | The prompt's content locale, e.g. en-US. |
Returns: Matching prompts, each with its score and the models it was tested against.
get_prompt
Reads one prompt.
| Argument | Required | Description |
|---|---|---|
| name | yes | Full path, e.g. support/refund-request. |
| version | no | A version label. Defaults to the production version. |
Returns: The rendered text, its variables, license and version.
get_evals
Reads a prompt's eval results.
| Argument | Required | Description |
|---|---|---|
| name | yes | Full path, e.g. support/refund-request. |
| version | no | A version label. Defaults to the production version. |
Returns: The eval cases and the score per model, with cost per 1,000 calls.
validate_prompt
Scores any system prompt against the rubric.
| Argument | Required | Description |
|---|---|---|
| text | yes | The system prompt to score. |
Returns: The same result /validate shows for that text.
get_prompt and get_evals only read public prompts, and a private one gets the same answer as one that doesn't exist. validate_prompt shares its rate limit with /validate; see protocol and limits.