Semble

Semble is an open-source code search library built specifically for AI coding agents. It lets agents like Claude Code, Cursor, Codex, OpenCode, and Gemini ask questions in natural language (e.g., "How is authentication handled?") and receive only the relevant code snippets — no grepping, no reading entire files. This replaces the traditional grep-and-read workflow with a token-efficient retrieval that can index an average repository in about 250 ms and answer queries in roughly 1.5 ms, all on CPU. Under the hood, Semble splits code into tree-sitter–aware chunks and scores them with a hybrid approach: static Model2Vec embeddings (using the code-specialized potion-code-16M model) for semantic similarity, plus BM25 for lexical matches on identifiers and APIs. The two lists are fused via Reciprocal Rank Fusion, then reranked with code-aware signals like definition boosts, identifier stem matching, file coherence, and noise penalties for test or legacy files. The result is retrieval quality on par with larger code-specialized transformer models — an NDCG@10 of 0.854 on benchmarks — at a fraction of the size and cost. Semble runs entirely on CPU with zero external dependencies. No API keys, GPU, or cloud services required. You can use it as an MCP server, a CLI tool via AGENTS.md, or a dedicated sub-agent. The interactive installer, `semble install`, detects your installed coding agents and lets you enable the integrations you want. It also supports indexing both local paths and git URLs, and offers two MCP tools: `search` and `find_related`. For developers and teams who want fast, private, and token-efficient code search without sending code to a cloud service, Semble is a lightweight local alternative to managed RAG setups. It's MIT-licensed and open source, so you can inspect, modify, and self-host it as you see fit.