The words behind the record.
Plain definitions of the terms InterfaceDelta uses, starting with MOSHCs, what most people call AI agents.
MOSHC
Machinic Object with Simulated Human Capacities; what most people call an AI agent
A machine that is technically made of computational parts but socially made by the purposes, authorities and relationships through which those parts act. A MOSHC can actually perform some human-associated work, such as drafting, scheduling, researching or negotiating. What is simulated is its human social presence, not the work itself.
The term comes from InterfaceDelta’s research on how machines become social actors. It is not a strict synonym for “AI agent”: it names the machine together with the role it is given in a particular setting.
Elsewhere on this site we use the everyday words, agent and tool, so that the record stays easy to find.
Agent tool
service, MCP server, endpoint
A service that an agent connects to in order to act: an MCP server, an OpenAPI or A2A endpoint, or an agent-commerce endpoint. One service can offer many tools.
Observation
What we saw when we looked at a service from the outside, at one moment: what it declares (its tools and their inputs, protocol details, payment terms) and whether it answered. We never call a service’s tools or send credentials.
Passport
The record of one observed service: its tool inventory, protocol details and payment terms, with the history of every observed version.
Change class
How a recorded change between two observations is classified, by fixed rules rather than AI.
Breaking: a change that can stop an existing caller working, for example a tool removed or an input made required.
Indeterminate: we could not tell, for example because part of the declaration could not be read.
Non-breaking: a change that should not affect an existing caller, for example a tool added.
Informational: a recorded change that does not alter what a caller can do, such as an edited description.
A first observation is the baseline. When the declaration format itself changed, two observations are recorded as not comparable.
Effect reading
AI reading
An AI classification of what a tool can do: for example reads only, creates, deletes, sends messages, moves money, runs code or changes permissions. Every reading is marked ◇ AI, carries a confidence, and never overrides a deterministic fact.
Facts, not verdicts
We publish what was observed and how it is classified. We do not rank, certify or call a service safe or unsafe, and nobody can pay for placement. The decision stays with you and your own policy.
Record Check
With a Record API token, one signed answer an agent can request before using a tool: whether it is observed, how old the observation is, whether it changed since a version you approved, its effect readings, and whether it asks for payment. Anyone can verify the signature.
Signed label
A compact statement of a service’s observed capabilities, signed with our published keys so that anyone can check it was issued by us and has not been altered.