How I built deferred tool discovery for my desktop AI assistant (no embeddings needed)

The developer of Ankita, an open-source desktop AI assistant, has implemented a deferred tool-discovery system to optimize token usage and performance. By avoiding the upfront loading of all tool schemas, the assistant saves context space, which is critical for maintaining efficiency in LLM-powered applications. Instead of using complex vector embeddings, the system utilizes a single 'find_tools' function that matches user queries against a curated, keyword-based catalog. This approach ensures predictability and debuggability while keeping the CLI free of runtime dependencies. Additionally, the assistant supports 'skills'—procedural instructions stored as markdown files—which are loaded on demand. This architecture allows the assistant to remain lightweight and responsive, demonstrating that simple, rule-based logic can often outperform more complex machine learning solutions for specific desktop automation tasks. The project is available on GitHub for community feedback and further development.