MCP tool reference
Sumi-Docs-MCP exposes four stateless tools. Every argument object is validated strictly; unknown fields and malformed values are rejected before execution.
| Tool | Arguments | Result |
|---|---|---|
list_docs |
{} |
Document paths, titles, optional timestamps, and optional public URLs |
search_docs |
{ "query": string } |
Ranked lexical matches with headings, snippets, and optional URLs |
fetch_doc |
{ "path": string } |
Parsed content, frontmatter, headings, and an optional public URL |
get_openapi_spec |
{ "endpoint"?: string } |
The full OpenAPI object or one exact endpoint |
Search queries must contain 1 to 200 characters. Search is Unicode-aware lexical substring matching, not semantic or embedding search. Equal scores are ordered by path so repeated calls over the same corpus are deterministic.
fetch_doc accepts a restricted relative .md or .mdx path already present
in the index. It never opens a client-supplied filesystem path.
Tool errors set isError: true and return a sanitized message. _meta.errorCode
is one of PATH_NOT_FOUND, INVALID_INPUT, or PARSE_ERROR; absolute paths and
stack traces are not exposed. Successful and failed results also identify the
MCP protocol version and server capabilities in _meta.
Stdio loads one read-only corpus snapshot on the first content tool call. Streamable HTTP loads the same snapshot before listening. Later calls reuse it until the process exits, so source changes require a restart.