Semble Launches Token-Efficient Code Search for AI Agents
Semble provides CPU-only, token-efficient code retrieval for agents via natural language queries and MCP or bash interfaces.
Semble returns exact code snippets using ~98% fewer tokens than grep+read, with full-repo indexing and search completing end-to-end in under one second on CPU (https://github.com/MinishLab/semble). It reports NDCG@10 of 0.854 on internal benchmarks while delivering ~200x faster indexing and ~10x faster queries than code-specialized transformers at 99% of retrieval quality.
The library exposes a search tool accepting natural-language or symbol queries plus a find_related tool that returns semantically similar chunks given a file path and line number. Local paths trigger automatic re-indexing on file changes; git URLs are cloned and cached for the session duration.
Integration occurs either as an MCP server for Claude Code, Cursor, Codex and OpenCode or via bash commands added to AGENTS.md files, requiring no API keys, GPUs or external services.
Claude: Semble removes the grep-and-read token bottleneck, enabling agents to maintain longer coherent reasoning traces across large repositories.
Sources (2)
- [1]Primary Source(https://github.com/MinishLab/semble)
- [2]Related Source(https://arxiv.org/abs/2103.05762)