May 17, 2026
Grep Wars: Token Edition
Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep
A flashy new code-finder says it crushes search costs, but the comments instantly yelled “wait, what?”
TLDR: Semble is a new tool that helps AI coding assistants find the right code faster while reading far less text. Commenters were split between curiosity and pure roast mode, with the biggest fight over whether comparing it to grep makes any sense at all.
A new tool called Semble arrived on Show HN promising a very online dream: help AI coding assistants find the right bits of code fast, cheaply, and without sending your laptop into orbit. The pitch is simple enough for non-experts: instead of making an AI read huge piles of files, Semble tries to grab just the exact snippets it needs. Its creator says it can search a whole project in a blink and use 98% fewer tokens than old-school search methods used by AI helpers.
But the real action was in the comments, where the crowd immediately went into cross-examination mode. The funniest jab came from the user who basically asked, “grep doesn’t need tokens, so 98% fewer than zero?” Ouch. That became the thread’s instant punchline, turning the launch into a mini courtroom drama over whether the comparison was clever marketing or just awkward wording. Others pushed for the test that really matters: not lab scores, but whether real AI coding tools like Claude Code or Copilot would actually trust the results instead of panic-searching the same files again and burning all the savings anyway.
Not everyone came to throw tomatoes. One commenter pointed out that semantic code search sounds useful for humans too, not just AI bots, while another wondered if it could also search things like API docs or AI memory notes. So the mood was classic Hacker News: one part impressed, one part deeply skeptical, and one part delighted to roast the benchmark claim into dust.
Key Points
- •Semble is presented as a CPU-only code search library for agents that can index and search full codebases in under a second.
- •The article claims benchmark results of about 200x faster indexing and 10x faster queries than a code-specialized transformer, with 99% of its retrieval quality.
- •Semble supports local directories and remote Git repositories, and returns matched code chunks with file path, line numbers, and content.
- •The tool can run as an MCP server for clients including Claude Code, Cursor, Codex, and OpenCode, with repositories cloned and indexed on demand.
- •Its retrieval pipeline combines Chonkie-based chunking, Model2Vec embeddings with potion-code-16M, BM25 lexical search, Reciprocal Rank Fusion, and code-aware reranking signals.