[FEATURE] MCP Profiles for Rapid Context-Aware Tool Switching

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

MCP configuration is currently all-or-nothing: configure an MCP server and it becomes available everywhere (Claude Code, Cowork, all projects). This creates three operational pain points:

1. Context waste from unintended MCPs: Developers configure MCPs for Cowork workflows (e.g., Slack for knowledge work) and those tools now consume 10-15% of Claude Code's context budget—even when Code doesn't need them. There's no way to disable them without removing them entirely.

1. No scoped tool management: disallowedTools doesn't work for MCP servers (#12863). Project-level exclusion doesn't work (#8321). Developers must manually edit .claude.json between sessions or maintain separate machine profiles—friction that breaks rapid task switching.

1. Behavioral drift: Without tool scoping, Claude Code may autonomously search/use MCPs it wasn't intended to access. Example: mentioning a person's name while working on a GitHub issue triggers Slack searches—unexpected autonomous behavior that requires explicit negative instructions to prevent.

Proposed Solution

Implement bundled MCP profiles that can be loaded, switched, and hotloaded per context:

{
  "mcp": {
    "profiles": {
      "coding": {
        "description": "Claude Code defaults",
        "mcpServers": {
          "github": { "..." : "..." },
          "context7": { "..." : "..." }
        }
      },
      "knowledge-work": {
        "description": "Cowork defaults",
        "mcpServers": {
          "slack": { "..." : "..." },
          "notion": { "..." : "..." },
          "github": { "..." : "..." }
        }
      },
      "empty": {
        "description": "No MCPs",
        "mcpServers": {}
      }
    },
    "defaults": {
      "code": "coding",
      "cowork": "knowledge-work",
      "project": "coding"
    }
  }
}

Scoping options:

• Tool-level: cowork profiles only, code profiles only, shared
• Session-level: claude --mcp-profile coding
• Project-level: .mcp.json or .claude/settings.json specifies profile override
• Runtime hotloading: /mcp-profile switch knowledge-work
• Default fallback: null or empty profile loads nothing

Profiles loadable from:

• User level: ~/.claude/mcp-profiles.json
• Project level: .claude/mcp-profiles.json
• Profile inheritance: Project overrides user; explicit flags override both

Alternative Solutions

Current workarounds, all unsatisfactory:

• Manually editing .claude.json between sessions to add/remove MCP servers
• Maintaining separate machine profiles for different workflows
• Setting ENABLE_CLAUDEAI_MCP_SERVERS=false and losing all MCP access in Claude Code
• Adding explicit negative instructions ("never use Slack MCP") to prevent autonomous behavior
• Using permissions.deny to block specific MCP tools (coarse, no bundling)

Priority

High - Significant impact on productivity

Feature Category

MCP server integration

Use Case Example

Platform engineer working on infrastructure:

1. Morning: claude --mcp-profile coding — loads GitHub, context7, Netbox for Terraform/Kubernetes work
2. Incident review: /mcp-profile switch knowledge-work — loads Slack, Notion, Linear for knowledge synthesis
3. Pair debugging with sensitive credentials: /mcp-profile switch empty — zero MCPs loaded

Today: Loses 10% context to Slack MCPs in Code sessions. Must manually edit config to exclude them. Claude autonomously searches Slack when names are mentioned in code context.

With profiles: MCPs load/unload atomically per context. Behavior is explicit and predictable.

Additional Context

Refiled from #24000 (auto-closed for inactivity, locked without team engagement) and freeform #39487.

This feature request subsumes:

• #7328: MCP tool filtering (granular)
• #7289: Disabling MCP in specific contexts
• #8321: Project-level MCP exclusion

Supporting data point from #24000: /doctor flags cloud MCP integrations consuming ~31K tokens of context overhead, with Notion alone at ~22K across 13 tools.