Semble, an open-source code-search engine built by Amsterdam-based MinishLab, hit the front page of Hacker News on Thursday after its creators claimed the tool consumes 98 % fewer tokens than the traditional grep-and-read workflow used by AI agents, according to github.com.
The project, posted by MinishLab co-founder Tom Aarsen, replaces line-by-line file reading with semantic embeddings that compress entire repositories into dense vector indexes. Aarsen explained in the thread that Semble pre-computes code embeddings once, then lets agents query the index with natural language or regex patterns without re-reading source files. The repository, licensed under Apache 2.0, already includes Python bindings and a pre-built Docker image for plug-and-play deployment.
For AI-coding startups, token efficiency translates directly to cost. A mid-sized repository that would burn 50,000 tokens per grep pass now needs roughly 1,000 tokens with Semble, according to benchmarks cited in the GitHub README. That saving is comparable to the 90 % reduction Anthropic advertised when it introduced Claude’s “prompt caching” last year. By open-sourcing the approach, MinishLab is positioning itself alongside other vector-search specialists such as Sourcegraph’s Cody and GitHub’s Copilot Workspace, both of which still rely on heavier retrieval pipelines.
The team plans to add support for 30 more programming languages by July and is inviting external contributors to extend the embedding model to documentation and issue trackers. Aarsen told HN commenters that a hosted cloud version with usage-based billing will enter closed beta next month; interested teams can join a wait-list linked from the repository’s landing page.