MinishLab released Semble — a code search library purpose-built for AI agents. It returns the exact code snippets an agent needs instead of having grep dump full files into the context window. The result: roughly 98% fewer tokens used per code-search step. Show HN hit 212 points this morning.
## The numbers
End-to-end indexing of a full codebase plus query: under 1 second. About 200× faster indexing and 10× faster queries than code-specialized transformers, while keeping 99% of retrieval quality. All on CPU — no GPU, no API keys, no external service. Drop-in replacement for grep+read in any agent loop.
## Natural-language interface
Instead of grepping for a keyword and reading the file, you ask: “How is authentication handled?” Semble returns just the relevant code snippets. For Claude Code, Cursor, Codex, OpenCode, and other agents, it ships as an MCP server — the agent automatically picks Semble whenever it needs to find code.
## Why it matters
The expensive thing in agentic coding isn’t model inference, it’s context — every grep+read burns tokens for code the agent doesn’t end up using. Semble removes 98% of that waste while staying CPU-only. If you’ve felt your coding agent slow down on a big repo, this is the missing piece.
Top AI Product 
Leave a comment