MCP server

Score, read, and author models from an AI client over the Model Context Protocol.

The U+22A8 MCP server exposes the platform to AI clients over the Model Context Protocol. An agent connected to it can read models, score and compare content, and author and train models — the same operations as the REST API, bound to the same namespaces, reached through tool calls instead of HTTP requests.

https://u22a8.ai/mcp

This is the product MCP server — it scores content. It is distinct from the documentation MCP that this site hosts for answering questions about the docs. When a client lists both, the product server is the one at u22a8.ai/mcp.

Authentication

The server authenticates with OAuth, not a static key. On first connection a client opens a browser sign-in; once authorized, the session reaches the same models a REST key would and respects the same namespace ownership — a session authors only in namespaces it owns and reads its own plus the public catalog (u22a8, live, bench).

Any MCP client that supports remote servers with OAuth discovers the authorization server and runs the flow automatically. No key is pasted into client configuration.

Connect

Point an MCP client at the server URL. Two common forms:

Claude Code

claude mcp add --transport http u22a8 https://u22a8.ai/mcp

The client prompts for sign-in the first time a tool is called. After that, ask the agent to score against a model from the catalog:

Score text from an AI client

CopiedCopy prompt

Tools

Tools are grouped by what they do. Reading and scoring tools work against any model a session can read; list_samples, list_versions, and get_version additionally require ownership of the model’s namespace; authoring and lifecycle tools always require ownership.

Reading

Tool	Purpose
list_models	List models in a namespace (defaults to the public catalog).
get_model	Fetch one model’s metadata.
list_traits	List a model’s traits with sample counts and polarity labels.
list_samples	List a model’s training samples, filtered by trait. Owner-only.
list_versions	List a model’s versions, newest first. Owner-only.
get_version	Fetch one version’s snapshot by vN number or tag. Owner-only.

Scoring

Tool	Purpose
score	Score content against a model’s traits. Accepts a feedback map to capture labels as new samples.
compare	Score two texts and return the per-trait delta and overall improvement.
extract	Fetch a URL and return its extracted text — the same text score would read.

Authoring

Authoring tools mutate a draft model. Trait and sample edits apply only while the model is in the draft state.

Tool	Purpose
create_model	Create a draft model in a namespace you own.
update_model	Patch model metadata.
delete_model	Archive a model.
add_traits	Add or replace traits.
update_trait	Edit one trait.
delete_trait	Remove one trait.
add_samples	Append labelled samples.
delete_sample	Remove one sample.

Lifecycle

Tool	Purpose
train	Train a draft model, or re-tune a ready one. Runs asynchronously.
discover_traits	Derive intrinsic traits from a draft model’s samples. Runs asynchronously.
synthesize_samples	Generate labelled samples from a draft model’s trait definitions. Runs asynchronously.
tag_version	Pin a tag to the active version.
untag_version	Remove a tag.
activate_version	Point the model at a version — roll back or pin.

train, discover_traits, and synthesize_samples return once the run is enqueued; the model reports a busy state until the run completes. Poll get_model for the state rather than waiting on the call.

Models as resources

Beyond tools, the server publishes each readable model as an MCP resource. resources/list enumerates the models a session can read; reading a model’s URI returns its model card — the same markdown summary GET /v1/models/{handle} returns with Accept: text/markdown, covering the model’s description, current version, traits, and any additional terms.

https://u22a8.ai/v1/models/{handle}

Read a model’s card before scoring against it to understand what each trait measures and on what terms.

Next

REST API

The same operations over HTTP.

Score card

What the score tools return.