NextPlaid / ColGREP
Local semantic code search combines regex filters with meaning-based ranking, backed by a self-hosted retrieval engine with CPU and CUDA support.
ColGREP helps developers and coding agents find relevant code even when they don't know the exact words to search for. It runs locally, while NextPlaid, the search engine behind it, can also serve other retrieval applications on a self-hosted server.
- Code search: Combine regex filters with semantic ranking to find code related to a concept, with an index that refreshes when searches detect file changes.
- Local search agent: Hand a question to a small on-device model that searches repeatedly, reads candidate code and returns the lines that answer it.
- Coding agent integrations: Claude Code has full hooks support; OpenCode, Codex and Hermes have basic integrations.
- Retrieval server: NextPlaid provides a REST API with built-in model encoding, document additions and deletions without a full rebuild, and metadata filters that narrow the search.
ColGREP runs on macOS and Linux as a standalone binary without a separate server or external inference API. Its code search keeps your code on your machine. Agent mode requires a one-time model download. The default search model is lightonai/LateOn-Code-edge; lightonai/LateOn-Code and compatible HuggingFace ColBERT-style models are alternatives, with separate indexes for each project and model.
NextPlaid runs in Docker and uses ONNX Runtime for encoding. Its token-level matching preserves detail beyond a single document embedding, while compressed, disk-backed indexes reduce RAM use. It's optimized for CPU execution and also supports CUDA. Both products are open source under Apache 2.0.
