Semble code search cuts token usage 98% for AI agents

Semble is a new code search library purpose-built for AI agents, offering massive efficiency gains over traditional grep-based approaches. Instead of combing through entire files with keywords, agents can query in natural language (e.g., „How is authentication handled?“) and receive only the exact code snippets needed. The system claims to use ~98% fewer tokens than the typical grep+read workflow, dramatically reducing both cost and latency. An average repository indexes in about 250 milliseconds, and searches complete in around 1.5 milliseconds, all running on a standard CPU.

Built for seamless agent integration, Semble can run as an MCP server or be invoked via bash. It supports tools like Claude Code, Cursor, Codex, OpenCode, and any MCP-compatible agent. The library indexes local directories or remote git URLs, caches indexes per session, and watches local paths for automatic re-indexing on changes. For sub-agents that cannot call MCP tools directly, a bash integration via AGENTS.md/CLAUDE.md is also provided. No API keys, GPUs, or external services are required.

Benchmarks show Semble achieves an NDCG@10 of 0.854, on par with code-specialized transformer models like CodeBERT, while being roughly 200x faster to index and 10x faster to query. This combination of speed, accuracy, and zero-setup makes it ideal for developers looking to reduce token costs and improve agent response times in code-oriented tasks. The tool also tracks token savings with a `semble savings` command.