[FEATURE] Code-only scope for claude.ai connectors, or defer connector tool names until used

Request

A way to keep a claude.ai connector out of Claude Code sessions only, while it stays on in claude.ai chat and Cowork. Or defer connector tool names, not just their schemas, until a session needs that connector.

Why

Tool search already defers schemas: /context on our setup shows about 1.5M tokens of MCP tool definitions as deferred. But every connector's tool names still load into every session. With a typical set of claude.ai connectors that is about 1,900 names, around 40K tokens (measured from a session transcript's deferred-tools attachment), at the start of every Code session, before the first prompt.

Today the per-connector options don't fit:

• The desktop + > Connectors switch is account-wide (it also disables the connector in chat and Cowork). See the companion bug.
• disableClaudeAiConnectors, ENABLE_CLAUDEAI_MCP_SERVERS and deniedMcpServers are documented to act only on connectors Claude Code fetches itself. The desktop app delivers connectors in-process, and the docs say "No MCP setting or managed-mcp.json reaches those connectors."
• The only lever that appears to work is a project permission deny rule with a tool-name glob (e.g. mcp__<connector>__*). That is a permissions feature used as a context control, and its effect on the deferred name list is not documented.

Proposal (either would do)

1. A per-surface connector scope: "on in chat/Cowork, off in Code", or a per-project allowlist for desktop-delivered connectors.
2. Defer names too: list a connector once (name and one-line description) and load its tool names only when the session first uses that connector, the way chat's "load tools when needed" behaves.

Environment

• Claude desktop app, Code tab, local session
• Claude Code 2.1.296
• macOS 27.0.1