MCP from A to Z · Episode 15 of 90
MCP capabilities: how both sides say what they support
In the 2026-07-28 MCP spec, clients declare capabilities on every request and servers declare theirs in server/discover. Here are the fields, the rules and the new extensions map.
On this page
Client meets server. How does each know what the other supports? Capabilities. Last episode, server/discover told us what a server offers. Today: the capability objects themselves.
Two sets
There are two capability objects: ClientCapabilities and ServerCapabilities.
Client capabilities: on every request
In stateless MCP, the client declares its capabilities in every request's _meta, under io.modelcontextprotocol/clientCapabilities. The schema puts it plainly: "Capabilities are declared per-request rather than once at initialization; an empty object means the client supports no optional capabilities."
The fields in 2026-07-28:
elicitation, withformandurlmodes.rootsandsampling: still in the schema, but now deprecated.extensionsandexperimental.
Server capabilities: in server/discover
The server declares its capabilities in the server/discover result:
tools,resources,promptsandcompletions.tools,promptsandresourcescarry alistChangedflag, andresourcesalso hassubscribe.loggingis still listed, but deprecated in this version.- Plus
extensionsandexperimental.
The rules
- No reliance: "A server MUST NOT rely on capabilities the client has not declared." It also must not infer capabilities from earlier requests.
- Missing client capability: if processing a request needs one the client didn't declare, the server returns
MissingRequiredClientCapabilityError, code -32021. - Missing server capability: call
prompts/liston a server that didn't advertiseprompts, and you get method not found, code -32601. - Elicitation modes: "Servers MUST NOT send elicitation requests with modes that are not supported by the client."
New: extensions
This version adds an extensions field to both sides: "a map of extension identifiers to per-extension settings objects." Extension IDs follow the _meta key naming rules, with a mandatory prefix, such as io.modelcontextprotocol/ui or io.modelcontextprotocol/tasks.
If only one side supports an extension, the side that does must either fall back to core behavior or reject the request with an error.
Next episode: Errors, and the difference between protocol errors and tool errors.
Declare what you support. Never assume the rest.