Reduced Context
Bellatrix One optimizes AI performance by intelligently managing tool context. Instead of sending full tool specifications with every chat prompt, only relevant tools are included when needed.
The Problem with Full Context
Traditional MCP implementations send complete tool specifications to the AI with every request:
- Hundreds or thousands of lines of tool definitions
- Detailed parameter schemas for every available tool
- Example values and documentation for all functions
- This information repeated with every single message
This creates several problems:
- Increased token usage and higher costs
- Slower response times due to larger prompts
- Context window filled with irrelevant tool definitions
- Reduced space for actual conversation and task context
- Lower quality responses as the AI struggles with information overload
How Reduced Context Works
Bellatrix One's MCP Broker intelligently filters tool definitions:
- The AI receives a lightweight list of available tool names and brief descriptions
- When the AI needs to use a specific tool, it requests the full specification
- Only the relevant tool details are loaded into context at that moment
- After use, the detailed specification is removed from context
- The next request starts fresh with minimal context overhead
Benefits
Lower Costs
By reducing token usage by up to 90% on tool definitions, you save significantly on API costs. This is especially important for teams with many tools or high message volumes.
Faster Responses
Smaller prompts mean faster processing by the AI model. The AI spends less time parsing irrelevant tool definitions and more time on your actual request.
Better Results
With more context window available for your conversation, the AI can maintain better understanding of your task. It's not distracted by dozens of unrelated tool specifications.
Scalability
You can connect hundreds of tools without degrading performance. Traditional approaches break down when you have more than a handful of tools.
Technical Implementation
The reduced context system works through the MCP Broker:
- Tool discovery returns only names and one-line descriptions
- Full schemas are fetched on-demand when tools are invoked
- Intelligent caching reduces redundant specification requests
- Natural language tool search helps the AI find relevant tools without seeing all definitions
Real-World Impact
For a typical setup with 50 available tools:
- Traditional approach: ~15,000 tokens per request for tool definitions
- Reduced context: ~500 tokens for tool list, ~300 tokens when a tool is used
- Result: 95% reduction in tool-related token usage