`Semble`, token-light code search for agents using **98% fewer tokens** than grep

CPU-only hybrid retrieval gives coding agents a cheaper first pass before grep and full-file reads. Worth testing on large repos where context waste slows every task.

• —Combines potion-code-16M static embeddings with BM25, RRF, and code-aware reranking; no API key, GPU, or external service required.
• —Benchmarked on about 1,250 query/document pairs across 63 repos and 19 languages; reported NDCG@10 is 0.854.
• —Indexing is about 250ms for a typical benchmark repo and queries run around 1.5ms on CPU; very large repos may take longer.
• —Ships an MCP server for Claude Code, Cursor, Codex, and OpenCode, so it can slot into existing agent workflows.