MCP

The MCP server exposes scoring as Model Context Protocol tools that any MCP-compatible client can invoke. This page documents the endpoint, transport, configuration, and the three tools available.

Endpoint and transport

Endpoint URLhttps://u22a8.ai/mcp
TransportHTTP (MCP streamable HTTP). No WebSocket or stdio transport is offered on the public server.
Session semanticsStateless. Each tool call is independent; no session or persistent connection is required.
AuthenticationNone required.
Rate limit10 requests per 10 seconds per client. Requests beyond this limit receive an MCP error.

Client configuration

Claude Code
$ claude mcp add --transport http qed https://u22a8.ai/mcp
mcp.json — any MCP-compatible client
{ "mcpServers": { "u22a8": { "type": "http", "url": "https://u22a8.ai/mcp" } } }

Tools

score
Score text or a URL against a model's traits. Returns per-trait scores with tier labels, confidence, headroom, and breaks, plus the composite. Optionally compares two inputs side by side.
Parameters
profile_handle Required. Model handle, e.g. u22a8.technical-writing.
content Required. Text to score, or a URL to fetch and score.
compare_to optional Baseline text or URL. When provided, the return value also includes baseline scores and a delta.
traits optional List of trait names to restrict scoring to. Omit to score all traits on the model.
Returns

An object matching the /p/{handle} REST response shape (scores, composite, confidence, headroom, input_tokens, detail, and optionally compare_to and delta). See the REST API for the full schema.

Failure modes

Returns {"error": "…"} when the model does not exist, a URL cannot be fetched or extracted, or a named trait is not on the model.

list_profiles
List the public models available for scoring. Equivalent to the GET /profiles REST endpoint.
Parameters

None.

Returns

An object with a profiles array. Each entry includes handle, description, and traits.

list_traits
Inspect the traits defined on a model. Equivalent to the GET /p/{handle}/traits REST endpoint.
Parameters
profile_handle Required. Model handle to inspect.
Returns

An object with a traits array. Each entry includes key, name, description, positive_label, negative_label, and sample counts.

Failure modes

Returns {"error": "…"} when the model does not exist.

Scores are approximate — not a substitute for human judgment.