Semble
An open-source code search library that lets AI coding agents find relevant code using far fewer tokens than grep.
🔗 Visit Semble
Description
AI coding assistants like Claude Code or Cursor typically find relevant code the brute-force way: grep for a keyword, then read whole files hoping the right snippet is in there. That approach burns a lot of tokens on irrelevant text. Semble is built to fix exactly this — it indexes a codebase once, then answers plain-language questions about where the relevant code lives, in about 1.5 milliseconds per query.
Semble is an open-source (5,700+ GitHub stars) code search library aimed at AI agents and agent-based IDEs. It indexes a repository in roughly 250ms, then retrieves relevant code snippets via natural-language queries with claimed 99% retrieval quality parity to specialized transformer-based code search models — while using around 98% fewer tokens than a typical grep-and-read workflow. It runs entirely locally with no API keys, no GPU, and no external service dependency, and ships both a CLI and a Python library, plus an MCP server so it can plug directly into Claude Code, Cursor, and similar tools.
💬 Our review
The short version: if you're building or using an AI coding agent and you're burning tokens (and money) on grep-then-read workflows, Semble is a free, local, genuinely clever fix — the tradeoff is it's a newer project, so you're trusting its retrieval-quality claims more than a decade-proven tool.
Against plain grep-based retrieval (what most agent harnesses default to), Semble's pitch is straightforward: same relevant results, far fewer tokens, because it indexes semantically instead of reading whole files. Against heavier embedding-based retrieval setups that need a GPU or an external vector database, Semble's local, no-GPU, no-API-key design is a meaningful simplicity win for anyone who doesn't want to stand up extra infrastructure just for code search. The main thing to watch is that "99% retrieval quality parity" is a claim from a young, fast-moving open-source project rather than an independently audited benchmark — worth validating on your own codebase before trusting it in a production agent pipeline.
💰 Pricing
📊 Global score
🤖 AI-enriched data
100% open source et gratuit, aucune clé API ni GPU requis
Pros
~98% moins de tokens que grep+read
Indexation ultra-rapide (~250ms) et requêtes ~1.5ms
100% local, pas de clé API ni GPU
Serveur MCP compatible Claude Code / Cursor
Cons
Projet jeune, claims de qualité pas encore indépendamment audités
Communauté plus petite que grep ou Sourcegraph
Nécessite une phase d'indexation initiale