[Feature Request] Add MCP server scoping to reduce context usage in main conversation

Feature Request: Load MCP Servers in Agents Only (Context Isolation)

Problem

MCP tool definitions consume significant context (64.7k tokens / 32.3% in my case) even when tools aren't being used. All MCP tools are loaded into the main conversation regardless of need.

Observation

We already work around this with custom agents (e.g., bugsnag-analyzer that handles BugSnag queries). But the MCP tools are still loaded in the main context, wasting tokens.

Proposed Solution

Allow MCP servers to be scoped to agents only:

{
  "mcpServers": {
    "linear": {
      "command": "...",
      "scope": "agent-only"
    }
  }
}

Or define MCP servers directly in agent configurations:

# .claude/agents/linear-agent.md
---
mcpServers: ["linear"]
---
Handle Linear issue queries and updates.

Benefits

• Main context stays clean — no token cost for unused tools
• MCP tools load only when relevant agent is spawned
• Natural separation of concerns (Linear agent, Figma agent, etc.)
• Aligns with existing agent architecture