semble

What it does

Semble is a code search library optimized for agents that returns exact code snippets with semantic understanding, using 98% fewer tokens than traditional grep-and-read approaches. It indexes codebases in ~250ms on CPU and answers queries in ~1.5ms, with no external services or API keys required. The library achieves 0.854 NDCG@10 retrieval quality, on par with code-specialized transformer models at a fraction of the size.

Who it's for

Agents and developers building semantic code search into Claude Code, Cursor, or other MCP-compatible tools. Useful for anyone indexing and searching large codebases efficiently or integrating AI-powered code analysis into existing workflows.

Common use cases

• Semantic code search across a codebase without grepping individual files or reading full file contents.
• Integrating code search into Claude Code or Cursor as an MCP server for instant agent access to relevant snippets.
• Finding similar code patterns and implementations using the find-related command.
• Searching documentation and configuration files alongside code with the --content flag.
• Reducing token usage in LLM prompts by returning only semantically relevant code chunks.

Setup pitfalls

• High risk classification: reads and writes to the filesystem — ensure it only indexes codebases you intend to expose.
• Respects .gitignore and .sembleignore files to filter indexed content; review these rules carefully to avoid unintended exclusions.
• MCP server mode requires separate configuration per agent (Claude Code, Cursor, etc.) — manual setup if semble install doesn't detect your agent.
• When searching git URLs, repositories are cloned on demand and cached locally; ensure sufficient disk space and network access.