mrsladoje/sweet-search is a local code search tool that helps coding agents search repositories and find related context. It builds a repository index and reflects changes made during work. The README explains that indexing and search happen on the device and that code is not sent to the cloud. FACTS does not record a public release version.
The tool is configured to be used together with Claude Code, Codex and OpenCode. It provides hybrid search that reranks search results along with its own grep tool, and introduces support for GPU indexing and incremental indexing. Indexing is the process of building search-ready data for a repository, and incremental indexing is the way of reflecting later changes. Installation and per-agent connection methods, the index structure and comparison materials are also included in the README. The project is written in JavaScript and uses the Apache-2.0 license.
The README states that it lowered cost from 5 to 33 percent on 200 SWE-bench instances. This is a result reported by the repository, and with only the given material it is difficult to judge whether the same saving can be expected in other environments. So it is clear that this is a local search tool that can be connected to multiple agents, but the cost reduction figures should be read strictly as claims of that evaluation.