MinishLab released Semble, a code search tool built specifically for AI agents. The pitch is simple: it uses 98% fewer tokens than grep when searching code.

That matters because tokens cost money. When an AI agent searches a codebase, tools like grep return entire files or large chunks of matching code. The agent has to process all of it, which burns through context windows and racks up API costs fast.

Semble tries to fix this by being smarter about what it returns. Instead of dumping full files, it sends back structured, compressed results that agents can actually work with. The idea is to give the agent enough information to understand the code without paying for thousands of tokens of boilerplate.

How it works

The tool indexes your codebase and builds a searchable representation that’s optimized for LLM consumption. When an agent queries it, Semble returns compact summaries and references instead of raw code.

Think of it as the difference between showing an agent every line in a file versus showing it “here’s where the auth logic lives, here are the three functions that matter, and here’s their signatures.”

MinishLab claims this approach cuts token usage by 98% compared to naive grep-based searches. They haven’t published detailed benchmarks yet, but the efficiency gains make sense in principle. Less data in, less data out, same level of understanding.

Who should care

If you’re building AI agents that work with code, this is worth looking at. The token savings could be real, especially if your agents search large codebases repeatedly.

It’s particularly relevant for autonomous coding tools, CI/CD agents, or anything that needs to navigate unfamiliar repos without human handholding. The faster and cheaper you can get an agent oriented in a codebase, the more useful it becomes.

What’s missing

The GitHub repo doesn’t include much documentation yet. There’s a README with basic usage, but no deep dive into the indexing strategy or comparison benchmarks. It’s also unclear how well this works with different languages or project structures.

The project is brand new, so rough edges are expected. But if you’re serious about using it, you’ll probably need to dig into the code to understand what it’s actually doing under the hood.

The bigger picture

Code search for agents is becoming its own category. As more companies build AI-powered dev tools, the infrastructure around them is getting specialized attention. Semble is one example. There are others working on similar problems, like optimizing context retrieval or building agent-friendly code graphs.

The common thread is that tools designed for humans don’t always work well for agents. grep is great if you’re a developer staring at a terminal. It’s less great if you’re an LLM trying to understand a 50,000-line monorepo without blowing your context budget.

Semble is still early, but it’s tackling a real problem. If the token savings hold up in practice, it could become a standard part of the agent toolchain. Worth watching.