Semble: A Code Search Tool for Agents That Cuts Token Usage by 98% Compared to grep

Semble is a code search library designed for AI agents that achieves about a 98% reduction in token usage and dramatically lower latency compared to the traditional grep + read approach.

Key features are as follows:

• Overwhelming speed: Indexing an entire codebase takes less than 1 second, and query speed is about 10x faster than code-specialized transformer models.
• Natural language-based search: Instead of keyword matching, it extracts exactly the code snippets you need through natural language questions like "How is authentication handled?"
• Low resource requirements: It can run on CPU alone, without a GPU or API key, and has no dependency on external services.
• Broad compatibility: It can be run as an MCP (Model Context Protocol) server or integrated via Bash, allowing immediate connection with various AI agents such as Claude Code, Cursor, and Codex.