The archive · Developer & Business Tools · Technical decision · 2024–2026
Semble bets static embeddings + BM25 can beat grep for coding agents, saving 99% tokens
MinishLab's Semble gives coding agents CPU code search — Model2Vec static embeddings plus BM25 match transformer quality at ~99% fewer tokens than grep+read.
Semble (MinishLab)
What the business is
Semble is an open-source code-search library and MCP server for coding agents (Claude Code, Cursor, Codex, OpenCode): it indexes a repo on demand and returns the exact relevant code snippets in milliseconds, running entirely on CPU with no API keys, GPU or external services.
How it started
Stephan Tulkens and Thomas van Dongen built MinishLab's Model2Vec in 2024, a method that distills sentence transformers into small static embeddings. Using their own Claude Code sessions on large codebases, they kept hitting the same wall: when the agent could not find something directly, it fell back to grep, read whole files or spawned subagents — burning tokens and still missing relevant code. Existing retrieval tools were either too slow to index on demand, needed API keys, or had poor quality, so in early 2026 they started building Semble.
What happened
Semble was open-sourced in spring 2026 (an initial Show HN ran 2026-04-26; the main one, 2026-05-17, drew 445 points and 151 comments). It combines potion-code-16M static embeddings with BM25, fuses them via RRF, and reranks with code-aware signals like definition boosts and test-file penalties. On their benchmark of ~1,250 query/document pairs across 63 repos and 19 languages it scored 0.854 NDCG@10 — about 99% of a 137M-parameter code transformer — at roughly 200x faster indexing, all on CPU. The README's telemetry shows the bet landing: 14.3k search calls and ~714.2M tokens saved as of September 2026.
No ending yet — it is still running.
Background
Semble is an open-source code-search library and MCP server built by MinishLab, the two-person research team behind Model2Vec. It answers natural-language queries against a codebase — "How is authentication handled?" — and returns the relevant snippets in milliseconds, entirely on CPU, with no API keys, GPU or external services. Its target users are coding agents: Claude Code, Cursor, Codex and OpenCode.
The bet is that agents waste enormous numbers of tokens on naive exploration. When Claude Code could not find something, it fell back to grep, read whole files or spawned subagents. Semble combines a 16M-parameter static embedding model (potion-code-16M) with BM25, fuses the rankings with Reciprocal Rank Fusion, and reranks with code-aware signals. On its benchmark of ~1,250 queries across 63 repos and 19 languages it reached 0.854 NDCG@10 — about 99% of a 137M-parameter transformer — at roughly 200x faster indexing, with ~99% fewer tokens than grep+read.
The launch landed on Hacker News on 2026-05-17 with 445 points and 151 comments, and the project kept compounding: the README's built-in savings telemetry recorded 14.3k search calls and ~714.2M tokens saved by September 2026. Still an open-source library without a paid product, Semble is a live demonstration that retrieval quality can be decoupled from model size.
What has to be true
- It targeted a measurable pain — token burn — rather than a vague feature, so the value proposition was a number (98–99% fewer tokens) that users could verify themselves.
- It inverted the AI playbook: instead of a larger model or a GPU-heavy retriever, it used a 16M-parameter static model plus BM25, making the product free to run and zero-config.
- Shipping as an MCP server meant instant distribution: one command wired Semble into every major coding agent rather than requiring a new workflow.
- The founders built the tool for their own workflow first, which is why the quality bar (does this actually find the code) beat the architecture hype.
What can be applied
Before buying a bigger model, measure what an agent pays to find code — a 16M-parameter static model plus BM25 beats a 137M transformer at a fraction of the cost when the job is retrieval.
Aftermath
As of September 2, 2026, Semble is live and growing as an MIT-licensed open-source project: an interactive installer, a CLI, a Python library and an MCP server for Claude Code, Cursor, Codex, OpenCode and VS Code, with incremental re-indexing and a code-aware ranking pipeline. Its own telemetry reports 14.3k search calls and roughly 714M tokens saved, heaviest in the seven days before this date. There is no disclosed funding or commercial pricing; MinishLab keeps publishing open-source models around static embeddings, and Semble remains a library rather than a revenue company.
Sources
- Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep
- MinishLab/semble: Fast and Accurate Code Search for Agents
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card