#151 perf: local members wastefully load fleet-mcp on startup

[Anonymous]

Originally created by: kumaakh

Problem

Local fleet members (e.g. odm-dev, fleet-dev) currently load the fleet MCP server as part of their Claude Code configuration. For members that are consumers of fleet work (doers/reviewers), fleet-mcp is unnecessary — it is the PM's tool, not the member's. Loading it wastes startup time and memory on every invocation.

Options to explore

1. --bare flag — launch Claude Code without loading any MCP servers; member only gets the tools it needs for its role
2. Explicit exclusion — configure member settings to skip fleet-mcp specifically (e.g. via settings.local.json disabledMcpServers or equivalent)
3. Don't register fleet-mcp at user level — move fleet-mcp registration out of ~/.claude/settings.json (user-level) into a PM-only profile or project-level config, so it only loads when the PM is running
4. Role-aware MCP loading — fleet server composes member settings that only include MCP servers relevant to the member's role (doer/reviewer get no fleet-mcp; PM gets it)

Notes

• This only affects local members — remote members connect via SSH and the PM dispatches into a fresh Claude Code process each time, so the cost is per-invocation
• The fix should be transparent to the doer/reviewer: they should not need fleet-mcp to do their work
• Any solution must not break the PM's ability to use fleet-mcp to dispatch work

Priority

Low — correctness issue, not a blocker. Worth fixing to reduce overhead on busy local machines running multiple members simultaneously.

[Anonymous]

Ticket changed by: kumaakh

• status: open --> closed