Feature request: restrict MCP servers to specific models via a models field

What feature would you like to see?

An optional models field on MCP server entries (stdio, http, and sse) that restricts a server to specific models.

Today every configured MCP server is loaded for every model. Users who mix models — for example a strong primary model and a cheaper or specialized secondary one — cannot keep tool-heavy or provider-specific MCP servers away from models that should not use them.

Proposed semantics:

• models accepts exact model aliases and prefix* wildcards (e.g. example-provider/*).
• A server with models loads only for sessions whose model matches; servers without models load for every model as before (fully backward compatible).
• Restricted servers are excluded when the session's model is unknown.
• Scoping is evaluated once at session start; mid-session model switches do not reload MCP servers.

Additional information

I searched existing issues and PRs for "MCP models", "restrict MCP", and "scope MCP" and found no prior request for per-model MCP scoping (the closest items, such as [tools].disabled filtering, are global rather than per-model).

I have a working implementation of this proposal (v1 engine, with tests and docs) that I am currently testing on my fork: https://github.com/kevingatera/kimi-code/pull/2 — happy to open it as a PR here once maintainers signal the approach is acceptable. Note the v2 engine's separate MCP stack would need a follow-up to honor the field.

Start by reviewing the v1 engine's MCP configuration and session-start handling, then inspect the existing tests and docs mentioned in the proposal. Compare the separate v2 engine MCP stack to determine the scope of the follow-up. Done means the models field supports exact aliases and prefix* wildcards, preserves unrestricted behavior, excludes unknown models, and is covered by tests and documentation.