Research infrastructure

MCP servers

The Lab publishes some of its research as Model Context Protocol servers, so that a coding agent can query the material directly rather than a person copying answers out of a document.

Servers in operation

Each server is token-gated. Access is issued per person to fellows and collaborators. Each server page states what the server contains and how to request a key.

pai-bok

The Production AI Body of Knowledge. Every entry takes a field that already solved a reliability problem and states what its solution becomes when the system runs on a model.

Nine tools, divided between deciding what to build and asking what the corpus contains. Serverless: no instance runs while nobody is asking.

Documentation →

How access works

There is no self-serve path. Keys are issued to a named person and revoked the same way, so that the Lab can see how the material is used and where it falls short.

Email [email protected] with the subject “MCP access,” stating who you are and what you are building. Other contact routes are on the Contact page.

Rate limits are shared across all callers rather than allocated per key. These are research services, not a production API.

One key per person

Issued and revoked individually. The Lab does not publish tokens on this site. A token presented as coming from here did not.