[FEATURE] Skip disconnected platform connectors from context injection
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Summary
Claude Code currently injects the full claude.ai connector catalog into every session as deferred tool names — including connectors that are explicitly not connected. There's no way to suppress this via local configuration (~/.claude/settings.json or .mcp.json).
Current behavior
Every session inherits ~15+ platform connector tool stubs (Slack, Notion, Gmail, Jira, Amplitude, Cloudflare, etc.) in the context window regardless of their connection status. These appear as deferred tool names and consume tokens on every turn, even when the user has no intention of using them.
Expected behavior
Only connectors with an active, authenticated connection should be injected into the session context. Disconnected connectors should be excluded entirely — or at minimum, users should be able to opt out via a config flag.
Impact
- ~460 tokens/turn saved per session from disconnected connector stubs
- At scale (just ~600 commands), this adds up: ~$0.08 on Sonnet 4.6, ~$0.41 on Opus 4.8 — small but unnecessary overhead
- More importantly, it reduces context noise: users debugging tool behavior see fewer irrelevant tool names in the context
Proposed Solution
- Local allowlist — let .mcp.json or settings.json declare "platformConnectors": ["clickup", "github"] to restrict injection to named connectors only
- Opt-out flag — a "injectPlatformConnectors": false setting in ~/.claude/settings.json
Alternative Solutions
None currently. Connectors connected at the account (or org) level are automatically injected into every session across claude.ai and Claude Code. Unused connector tool schemas consume context window tokens on every request.
Priority
Medium - Would be very helpful
Feature Category
MCP server integration
Use Case Example
_No response_
Additional Context
_No response_
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗