How It Works

Understanding the Intelligent MCP Orchestrator

The Cognitive Dispatcher

The MCP Orchestrator acts as an intelligent router between your tasks and the best available tools. Instead of hardcoding which tool to use, you describe what you need (capabilities), and the orchestrator finds the optimal match.

User Request Tool Selection Execution | | | v v v +---------+ +-------------------------------------+ +-------------+ | Task |--->| 1. Parse capabilities required |--->| Execute | | "Write | | 2. Find matching tools | | via best | | code" | | 3. Score by capability match | | provider | | | | 4. Apply learning boost (M5) | | | | | | 5. Select highest scoring tool | | | +---------+ +-------------------------------------+ +-------------+ | v +-------------+ | Response | | + Metrics | | + Feedback | +-------------+

Key Concepts

Capability-Based Routing: Tools are matched by capability tags (e.g., "code_generation", "calculation"), not exact names. This decouples tasks from implementations.

Claude-First Bias (D-M1-003): External provider preference weighted toward Anthropic Claude with a 1.5x multiplier.

Internal Tools First (D-M1-004): Local MCP tools are tried before external providers for lower latency and no costs.

Learning Layer: The system remembers what worked and adjusts tool scores over time (0.5x to 2.0x boost).

Explore the Details

Tool Scoring - See exactly how tools are scored and ranked

Learning Loop - Understand the M3 to M5 to M1 feedback cycle

Example Flow - Walk through a real task step by step