Dynamic Tool Loading

Fp Switchboard offers 996+ tools across 32 services. Loading all of them into an AI session at once consumes 40-60K tokens of context window — often causing errors. Dynamic tool loading solves this.

The Problem

When connecting to the unified endpoint (/mcp/unified), all tools load at connection time:

Mode	Tools Loaded	Context Usage	Risk
Standard	~996	~40-60K tokens	Compaction errors, degraded performance
Dynamic	3 meta-tools	~1.5K tokens	None

Caution

Claude Desktop and Claude.ai can hit “maximum compaction” errors when loading 996+ tools. Use dynamic loading or custom bundles to avoid this.

How It Works

Add ?dynamic=true to your unified endpoint URL:

https://api.switchboard.fpdigital.ai/mcp/unified?dynamic=true

Instead of 996+ tool schemas, your AI receives 3 lightweight meta-tools:

Meta-Tools

Tool	Purpose	Returns
unified_list_services	List available services	Service names, tool counts, connection status
unified_get_tools	Get tools for one service	Full tool schemas for the requested service
unified_call_tool	Execute any tool	Proxies the call to the actual service tool

Typical Flow

1. AI connects and receives 3 meta-tools (~1.5K tokens)
2. User asks: “Check my Gmail inbox”
3. AI calls unified_list_services to see Google is connected
4. AI calls unified_get_tools with service: "google" to get Gmail tools
5. AI calls unified_call_tool with tool: "gmail_search" and the parameters

Claude.ai Web: Deferred Loading

Claude.ai web connections automatically use deferred loading — a complementary approach:

• All tool names are sent, but schemas are deferred (defer_loading: true)
• Claude fetches full schemas on-demand via tools/get when it needs them
• Works with Claude’s built-in Tool Search feature
• ~95% token reduction compared to standard loading

Note

Deferred loading is automatic for Claude.ai web connections (MCP OAuth tokens). No configuration needed.

When to Use Each Mode

Scenario	Recommendation
Claude Desktop with all services	?dynamic=true on unified endpoint
Claude.ai web	Automatic deferred loading (just connect)
Single service (e.g., just Gmail)	Per-service endpoint: /mcp/google
Curated tool set	Custom Bundle with selected tools
Cursor / VS Code	?dynamic=true or per-service endpoints

Configuration

Claude Desktop

{
  "mcpServers": {
    "fp-switchboard": {
      "command": "npx",
      "args": ["-y", "@anthropic-ai/mcp-remote",
        "https://api.switchboard.fpdigital.ai/mcp/unified?dynamic=true"],
      "env": {
        "MCP_HEADERS": "authorization:Bearer fps_unified_YOUR_TOKEN"
      }
    }
  }
}

Claude Code / Cursor

Use the same URL with ?dynamic=true in your MCP server configuration.

Comparison

Feature	Standard	Dynamic	Deferred	Bundle
Token usage	~60K	~1.5K	~3K	Varies
All tools available	Yes	Yes (on demand)	Yes (on demand)	Selected only
Extra API calls	0	2-3 per task	1 per tool	0
Best for	Small tool sets	Full access, low context	Claude.ai web	Focused workflows