Semble

If you build or use AI coding agents, Semble addresses a real pain: the cost and latency of reading entire files. By returning only relevant chunks, it slashes token usage by roughly 98% compared to grep+read workflows. The hybrid static-embedding plus BM25 approach delivers accuracy on par with transformer-based code search—NDCG@10 of 0.854—without the GPU requirement. We'd reach for this when working with Claude Code or Cursor on a multi-file codebase; the MCP integration is especially slick, giving agents two tools (search and find_related) that feel natural. Where Semble stumbles, it's largely by design: it's a local, library-level solution. There's no cloud dashboard, no team permissions, no real-time reindexing as files change. If you need shared search across a large organization or frequent incremental updates, you'll need to build those layers yourself. Compared to alternatives like CodeBERT or GPT-based code search, Semble is dramatically faster and cheaper at query time, but it won't match the depth of understanding from a large model for highly ambiguous queries. The open-source, MIT-licensed nature is a major plus for privacy and customisation. Best for solo developers and small teams who want instant, offline code navigation; less suited for enterprises seeking a managed service.