All chapters
  1. Start hereA protocol standardizes the call. It does not grant the permission
  2. Foundations · 1The primitive: a function, its arguments, its result, its side effects
  3. Foundations · 2Function calling: the model asks, your application executes
  4. Foundations · 3Client and server: process boundaries, JSON-RPC, errors and trust
  5. Foundations · 4What MCP is: host, client, server and what the protocol promises
  6. Protocol · 5Three surfaces: tools act, resources inform, prompts guide
  7. Protocol · 6A first MCP server: typed input, deterministic data, the raw exchange
  8. Protocol · 7Java and the SDKs: typed DTOs, validation and interoperability
  9. Protocol · 8Local or remote: stdio, HTTP, stateless requests and availability
  10. Platform · 9The database decision: direct SQL, a service boundary or an MCP adapter
  11. Platform · 10Sharing across teams: ownership, catalogs, contracts and versions
  12. Platform · 11Who gets access: users, agents, delegation and least privilege
  13. Platform · 12Enterprise authorization: tokens, audiences, gateways and no passthrough
  14. Platform · 13The threat model: injection, poisoned tools, confused deputies and egress
  15. Platform · 14Governance: who may publish, who approves, who answers the page
  16. Operations · 15Observability and audit: what was called, by whom, and why it was allowed
  17. Operations · 16Resilience: timeouts, retries, idempotency and overload at a shared server
  18. Operations · 17Testing and evaluation: contracts, authorization matrices and injection
  19. Reference · 18Advanced MCP: long tasks, large catalogs, tenants and evolving versions
  20. Reference · 19The ecosystem: SDKs, registries, gateways, identity and telemetry
  21. Reference · 20Enterprise reference architecture: one assessment platform, four teams
  22. Reference · 21Decision playbooks: function, API or MCP, and the records that explain why
  23. ReferenceGlossary
  24. AboutAbout this site and its sources
Reference

Glossary

Every term this site uses, each with what it is, what it is not, and how it shows up in the insurance example.

Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, what it is, what it is not, and how it shows up in the insurance example.

Two sentences hold the words apart. A protocol standardizes the call; it does not grant the permission. And: the model recommends; a person decides. A tool call is a request written by a model. Discovery is what a server chooses to show. Authorization is what the server decides when the call arrives, for this principal, on this data, now.

A2AA protocol for agents built on different frameworks to discover each other's capabilities and work together on tasks; complementary to MCP.ABACAccess control that grants or denies a request from attributes of the caller, the resource and the situation.Access tokenA string representing an authorization issued to a client, sent as a bearer token on every request to the resource it was issued for.AdapterCode that exposes a domain service through one interface, such as a REST controller or an MCP server, without owning its rules.ADRA short document that states one decision, its context, the options considered and its consequences, kept with the code.API gatewayA reverse proxy that applies cross-cutting policy (authentication, routing, rate limits, logging) to HTTP APIs before requests reach the services behind it.ArgumentA value a caller gives a function when it calls it.AudienceThe claim in a token (aud) that names the resource it was issued for; a server accepts only tokens that name it.Audit eventA durable record, for people, of who did what and why it was allowed or refused.Authorization decisionThe recorded result of deciding whether this principal may do this action on this data now: result, reason, deciding check and detail.Authorization matrixEvery principal against every tool, each cell with the answer the server must give, run as one test per cell.Authorization serverThe OAuth party that authenticates the person or workload and issues access tokens under the organization's policy.Bean ValidationThe Java standard that declares constraints as annotations on fields, parameters and return values, checked by a validator.Bounded retryA retry with a maximum number of attempts, a growing wait between them, and a deadline over all of them.CapabilitySomething useful a system can do for a caller, independent of how it is exposed.Capability negotiationEach side declares which protocol features it supports, and both keep to what the other declared.Change advisoryA review of a proposed change by the roles it affects, before deployment, scaled to the change's risk.Circuit breakerA switch in front of a dependency that stops sending calls after repeated failures and lets one call through after a cool-down.Community extensionThis site's label for an optional addition maintained by an open project or community outside the normative MCP specification; it promises only what its maintainers state.Compatibility testA test that runs a server against more than one client or protocol revision and checks each pairing works or fails with a clear answer.Confused deputyA program with authority that is tricked into using it for someone who lacks that authority.Connection poolA set of open database connections owned by one process and lent to one request at a time.ConsentA person's (or organization's) recorded agreement that a specific client may act for them with specific scopes.Consumer teamA team that runs its own application on top of a shared capability and answers for how that application uses it.Contract testA test that checks the shape of what a server promises on the wire, not the business result.Control modelWhich actor decides when an MCP surface is used: the model for tools, the host application for resources, the person for prompts.Control planeThe components that decide and configure what may happen: scopes, catalog states, token policy, limits; changed rarely and through review.Correlation idAn identifier for one end-to-end request, carried unchanged across every hop.Data planeThe path that carries the actual work on every request: host, gateway, MCP server, service and database, and the reply.Database accountAn identity the database itself knows, with its own permissions, separate from any person or agent.Decision playbookAn ordered list of questions that leads a team to a choice and says what to check before acting on it.Delegated accessA client acts for a person with permission the person granted, limited to what that client needs, and the token names both.DeprecationAnnouncing that a tool version will stop working on a named date, naming its replacement, and keeping it working and self-explaining until then.Domain data ownerThe team that owns the data and business rules behind a capability, and therefore what every field means.Domain serviceThe code that owns the business rules and the data of one domain; every interface to it is an adapter.DTOA plain structure that carries data between layers and has no behavior of its own.Dynamic catalogA catalog whose tools are published, changed, deprecated or retired while consumers are connected, so a caller's list can change after it was approved.Egress controlDeciding outside the model where outbound traffic from a host or server may go.Enterprise-managed authorizationAn official MCP extension that lets an organization control MCP server access centrally through its existing identity provider.Execution adapterThe application code that turns a model's tool call into a function call or a refusal.ExfiltrationData leaving the place it is allowed to be, often through a tool call or URL the model writes.FallbackWhat the host does instead of a call that failed or was refused, so the user still gets an honest answer.Fitness checkAn automated test that verifies an architecture decision still holds, run on every change.FunctionOrdinary code with arguments and a result, and possibly a side effect.Golden scenarioA named situation with a recorded expected result that every run must reproduce.Idempotency keyA key the caller generates once per intended write and sends with every retry, so the service can store the write once.isErrorThe field of a tool result that marks a failed execution, so the model can read why and correct itself.Java recordA Java class declared by its components, for which Java derives the constructor, accessors and equals; its compact constructor can reject bad values.JSON SchemaA standard JSON document that describes which fields a JSON value must have and what each may contain.JSON-RPC 2.0A small, transport-agnostic standard for request and response messages in JSON: method, params, id, result, error.Least privilegeThe principle that each caller holds the minimum permissions its job needs, and no more.Local MCP serverAn MCP server that runs on the host's own machine, started and stopped by it, usually as a child process over stdio.Long-running taskAn operation whose result arrives long after the request that started it, so the caller receives a durable handle and comes back for the result.MCP clientA connector inside a host that talks to exactly one MCP server.MCP extensionAn optional, named addition to the Model Context Protocol that both sides must declare before either may use it.MCP gatewayA proxy in front of one or more MCP servers that applies shared checks such as token validation, routing, rate limits and logging.MCP hostThe AI application a person works in: it holds the conversation, talks to the model, connects to servers and asks for consent.MCP Java SDKThe official Java library for MCP servers and clients, maintained in collaboration with Spring AI; Tier 2 on the MCP SDK page.MCP promptA message template an MCP server offers for a person to pick, filled in with arguments.MCP RegistryThe official public index of metadata about publicly accessible MCP servers, in preview as documented on 2026-11-18.MCP resourceData an MCP server offers, identified by a URI, that the host decides when to bring into context.MCP SDKA library that implements MCP's messages, metadata and transports so a server or client can be written as ordinary code.MCP serverA program that offers tools, resources and prompts to MCP clients, as a local process or a remote service.MCP server maintainerThe team that writes and releases an MCP server: tool names, schemas, descriptions, versions and the server's authorization code.MCP toolA capability an MCP server offers for the model to propose: listed with tools/list, run with tools/call.MCPAn open standard for how an AI application discovers and uses capabilities (tools, resources, prompts) that other programs offer.Mock clientA scripted MCP client used in tests: it sends exact messages to the real server and records every reply, with no model behind it.Multi round-trip requestThe 2026-07-28 pattern in which a server that needs more input returns an input-required result and the client sends the request again with the answers.NamespaceA name prefix tied to an owner who has proven control of it, so two owners cannot claim the same name.OAuthA framework in which a client obtains a limited access token from an authorization server to call a protected resource.Official SDKA library maintained in the MCP project's own organization that implements the protocol for one language, ranked in tiers by conformance and maintenance commitment.OpenID ConnectAn identity layer on top of OAuth 2.0 whose ID token tells a client application who signed in and how.OpenTelemetryAn open standard and set of tools for traces, metrics and logs; its semantic conventions agree on span and attribute names, including for MCP.Parameterized querySQL whose values travel separately from the statement text, as placeholders the driver fills in.Partial failureA failure where one side of a remote call did some or all of its work and the other side does not know it.Platform teamThe team that runs the shared infrastructure through which capabilities are offered: gateway, catalog, identity integration, shared metrics.Policy decision pointThe component that takes a request and the applicable policy and computes whether access is allowed or denied.Policy enforcement pointThe component that stands in front of a resource and lets a request through or stops it, based on a policy decision.PrincipalThe identity a request is made by or for: a person, an agent (a workload) or a service.Process boundaryThe line between two running programs that do not share memory, so a call must travel as bytes.Prompt injectionText written to make a model follow the writer's instructions instead of its operator's, typed directly or hidden in content the model reads.Protected resource metadataA JSON document in which a protected resource states its identifier and which authorization servers issue tokens for it.Protocol revisionA dated version of the MCP specification; the date marks the last backwards-incompatible change.Publishing lifecycleThe states a shared tool passes through from proposal to removal, with the one role that may move it at each step.PydanticA Python library that validates data against classes declared with type hints and produces a JSON Schema from them.RACI matrixA table of activities by roles that marks who is Responsible, Accountable (exactly one), Consulted and Informed for each.Rate limitA cap on how many calls one consumer may make in a period, so one burst cannot spend capacity others need.RBACAccess control in which permissions are attached to roles and principals receive roles.Read operationAn operation that returns information and changes no stored state.RecommendationThe agent's proposed outcome for a submission (approve, decline, request_info or escalate) with reasons and rule ids; never a decision.Reference architectureA documented design for a class of systems that names its components, their owners and responsibilities, the flows between them and the decisions behind it.Release couplingHow far one team cannot release without another team releasing too.Remote MCP serverAn MCP server that runs as a network service, operated by a named team, that many hosts reach at one URL over Streamable HTTP.RepositoryThe one class that holds all SQL of a domain, so other code asks for data by meaning instead of by table.Request idA value the client picks for each request, which the server copies into the response so the two can be matched.Request metadataThe _meta object in every request's params that carries the protocol version and the client's capabilities.Resource indicatorA parameter, resource, in which the client tells the authorization server which protected resource a token is for.Resource URIThe identifier of an MCP resource, written under RFC 3986; a template is a parameterized URI under RFC 6570.Return valueThe value a function hands back to its caller when it finishes.Risk tierA label on a catalog record that says how much harm misuse of a tool could cause, and so how much review, testing and operational care it needs.Rug pullA server changing a tool's definition or behavior after it was approved, while hosts keep using it under the old approval.Sanitized argumentsA copy of a tool call's arguments with secrets removed and free text replaced by its size.Schema generationProducing a tool's JSON Schema from code, such as a Pydantic model, type hints or a Standard Schema object, instead of writing it by hand.ScopeA named permission carried in an access token that says what the token may be used for.Semantic versioningVersion numbers in three parts, MAJOR.MINOR.PATCH, where each part promises something about compatibility.SerializationWriting a value as bytes that another program can read and rebuild, and reading it back.Server spoofingAn endpoint presenting itself as a server it is not, so a host sends it requests, data or tokens meant for the real one.Server-side request forgeryTricking a program inside a network into fetching a URL an attacker chose, to reach what the attacker cannot.server/discoverThe method every MCP server must implement that returns its supported versions, capabilities, instructions and identity.Service contractEverything a consumer may rely on about a capability, written down and owned: inputs, outputs, errors, meaning, access, limits and change policy.Side effectA change a function makes outside itself that stays after the call, such as a stored row.SpanOne timed piece of work inside one program, with a parent, a status and attributes.Spring BootA Java framework that wires an application's controllers, services and repositories and runs cross-cutting features such as method validation.Stateless requestA request that carries everything the server needs to process it, so any instance can answer it and no earlier request is assumed.stdio transportThe MCP transport in which the client launches the server as a child process and they exchange newline-delimited JSON-RPC on its standard input and output.Streamable HTTPThe MCP transport in which the server is an independent HTTP service with one endpoint and every client message is its own POST.structuredContentThe field of a tool result that carries the answer as JSON data, checked against the tool's outputSchema.SubmissionOne quote request as a structured JSON record, written once by intake from the broker's email.Tenant isolationKeeping each group's data invisible to other groups that share the same system, without revealing that the data exists.Timeout and deadlineA timeout is the longest one attempt may take; a deadline is the longest the whole call may take, all attempts included.Token exchangeA request in which a service trades a token it holds for a new token with a different audience, and usually less authority, issued under the authorization server's policy.Token passthroughAn MCP server forwarding the token it received from the client to another service; the MCP specification forbids it.ToolA capability offered to a model: the model requests it, application code executes it.Tool annotationAn optional hint in a tool definition about its behavior, such as read-only, destructive, idempotent or open-world.Tool callThe structured request a model writes to use a tool: an id, a tool name and arguments.Tool catalogThe organization's list of approved capabilities, one record per tool version, with owner, scopes, risk tier and lifecycle state.Tool definitionWhat the model is shown about a tool: its name, a description and a JSON Schema for its arguments.Tool poisoningPutting instructions or false claims into a tool's definition (description, schema, annotations) so the model or host misbehaves.Tool resultThe answer the application sends back to the model for one tool call, success or error.Tool-list filteringNarrowing the list of tools a caller is shown to the ones it may and should use, by identity, client and task.tools/callThe MCP request that asks a server to run one tool with arguments.tools/listThe MCP request that returns the tools a caller may see, with their names, schemas and annotations.TraceThe record of one client request as it passes through every program, made of timed spans.Trace evaluationAsserting on the recorded sequence of calls, spans and audit events of a run, not only on its final result.Transaction boundaryThe place in the code where a transaction starts and ends, and so who is responsible for consistency.TransportThe mechanism that carries messages between two processes, such as a pipe or an HTTP connection.Trust boundaryA point where requests or data arrive from a party the receiver does not control, so everything crossing it must be checked.UnderwriterThe employee who decides whether the insurer accepts a risk; the only principal who may record a decision.Underwriting agentThe example's agent: it gathers evidence with tools, applies deterministic rules and writes a recommendation for a person.Unit of workThe code object that opens one transaction for one business operation and commits or rolls it back.User identityAn identity that belongs to a person who signs in, holds a role and can be held responsible for a decision.Vendor featureThis site's label for a capability of one company's product, documented and changed by that company.Version evolutionThe managed change of a capability's contract over time on three clocks at once: tool versions, protocol revisions and extension revisions.Workload identityAn identity that belongs to a program, such as an agent or a batch job, registered by the team that operates it.Write operationAn operation that changes stored state, so the system differs after the call.

Each term is explained in full on one page and used everywhere else. Statements about the Model Context Protocol, named products and standards were read in their own documentation on 2026-11-18 (some on 2026-11-17); the sources are listed on About. For the decisions these terms support, see Playbooks.