Skip to content

Appendix C: glossary

Approval. A person's yes or no to a write, asked by the server before the handler runs (chapter 6).

Approver. A function that answers approvals when the client cannot ask a person: a ticket queue, a chat bot, a test.

Budget. The token limits for definitions and results (chapter 3).

Concise / detailed. The two response formats. Detail fields are only in detailed results (chapter 4).

Cursor. A signed, expiring token for the next page of a list, bound to the tool, caller and query.

Deadline. The time a call may take before it is cancelled (default 30 s).

Discovery. Whether the host gets every tool definition (direct) or three meta-tools to search them (search) (chapter 12).

Effect. What a tool does to the world: READ, WRITE or DESTRUCTIVE.

Handle. A signed reference to a stored result the model reads in pages with handles.read (chapter 8).

Host. The AI application that connects to MCP servers and talks to the model.

Idempotency. Running a write at most once per caller and arguments; repeats replay the first result.

Identity. The verified caller (ctx.identity): subject, tenant, scopes. It comes from the token, never from arguments.

Job. Background work started by a tool and polled with jobs.get; named jobs restart after a crash.

Lint. A check by ouncemcp check for definitions a model will misuse.

Policy. A rule that decides whether a caller may see and call a tool (chapter 7).

Preview. The text a person sees when asked to approve a write.

Resource. Data the host can attach to a conversation (chapter 9).

Shaping. Trimming a result to its budget while keeping it valid against its schema.

Startup context. The tokens of every definition the host sends at the start of a session.

Store. Where idempotency records, receipts, handles and jobs live; shared across replicas in production.

Tenant. The group a caller belongs to (a branch, a company); data never crosses tenants.

Trace. A record of one tool call: tokens, latency, outcome.