Repository navigation
Replies: 1 comment
|
MCP is now stateless, that means that all state must be carried over the wire, so a client chafing capabilities between requests is fine, but they can expect that might have an impact on responses. So it should only be done for known reasons, certainly not arbitrarily. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Title: Should a client declare the same capabilities on tools/list as on tools/call (2026-07-28)?
Question
In 2026-07-28, client capabilities travel on every request in
params._meta["io.modelcontextprotocol/clientCapabilities"]rather than once atinitialize. That lets a server shape its responses per request, includingtools/list.But the spec doesn't seem to say anything about consistency between requests. In particular, can a client list tools while declaring one set of capabilities, then call those tools declaring a different set? If it can, a server has no reliable way to give the model a tool schema that matches how the call will actually be handled.
Is there a requirement or recommendation here, or should there be?
Why it matters
Take a server with a tool that needs user confirmation before it acts:
elicitation: the server confirms through an elicitation dialog, so the tool's schema needs nothing extra.A server that does this returns a different
tools/listdepending on whether that request declareselicitation. That seems to be exactly the kind of per-request adaptation the envelope makes possible. It only works if what the client declares when listing matches what it declares when calling.What we observed from one client (Claude.ai, 2026-11-13)
We connected the same MCP server from a Claude.ai Team plan account and a Free plan account, and logged what each request declared:
tools/liston connectAnthropic/Toolbox1.0.0{}{}tools/listfrom the chatAnthropic/ClaudeAI1.0.0uiextension onlyuiextension onlytools/callfrom the same chatAnthropic/ClaudeAI1.0.0elicitation+uielicitation+uitools/listfrom "Refresh tools list"claude-ai0.1.0elicitation+uiuiextension onlyTwo inconsistencies stand out:
tools/listfrom connect or chat omitselicitation, while everytools/calldeclares it.elicitationon Team but not on Free. So a server that adapts its tool list shows different schemas on the two plans after a refresh, while the tool calls on both plans declare the same capabilities. The schema a user sees, and possibly the one the model works from, depends on the plan and on whether they've clicked refresh, not on how the calls are actually handled.These are trimmed captures from the same connection (
_metaonly):{"method":"tools/list","params":{"_meta":{ "io.modelcontextprotocol/protocolVersion":"2026-07-28", "io.modelcontextprotocol/clientInfo":{"name":"Anthropic/ClaudeAI","version":"1.0.0"}, "io.modelcontextprotocol/clientCapabilities":{"extensions":{"io.modelcontextprotocol/ui":{}}}}}}{"method":"tools/call","params":{"_meta":{ "io.modelcontextprotocol/protocolVersion":"2026-07-28", "io.modelcontextprotocol/clientInfo":{"name":"Anthropic/ClaudeAI","version":"1.0.0"}, "io.modelcontextprotocol/clientCapabilities":{"elicitation":{"form":{},"url":{}}, "extensions":{"io.modelcontextprotocol/ui":{}}}}}}The result is that the model can hold a schema and instructions written for a client without elicitation, while every call it makes declares elicitation and is handled that way. A server that validates strictly against the schema for the capabilities declared on the call may then reject arguments the model was told to send.
Possible directions
tools/listas on anytools/callthat may follow, or at least a superset of what a later call will declare.tools/list.Has anyone else hit this? Is there existing guidance we've missed?
All reactions