xerj: a cheaper way for AI agents to search

What it is

When an AI agent works on a project, it spends a lot of time hunting for things: where a phrase appears, which file mentions a client, where a setting lives. The usual tool for this is called grep. Think of it as flipping through every page of a binder looking for one word. It works, but the agent ends up reading plenty of pages it didn't need.

xerj is a new search engine made for exactly this job. It turned up in a roundup of trending new GitHub projects, described in a single line: compatible with Elasticsearch, and using 40x fewer tokens than grep for AI agents.

Why it caught our eye

Agents cost money based on how much they read. Less reading means a smaller bill and, often, a quicker answer. And because xerj speaks the same language as Elasticsearch, a well-known search engine many businesses already run behind the scenes, it might fit into what's already there instead of asking you to start over.

A quiet caveat: that one line is all we have. We don't yet know how the 40x was measured, what the license allows, or how fiddly it is to set up. Worth a look, not yet worth a bet.

Words worth knowing

• Tokens: the small chunks of text an AI reads and writes. You pay per token, so fewer means cheaper.
• grep: an old, trusty tool that searches files for exact words. Simple, but it makes agents read a lot.
• Elasticsearch-compatible: it understands the same requests as Elasticsearch, so tools built for one can often talk to the other.

Something to take with you

If a developer or agency runs AI agents for you, ask them one thing: how much of our bill is just the agent looking for stuff? The answer is a good conversation starter, and xerj is one place to begin: https://github.com/xerj-org/xerj