Summary

Local-first code search your editor can query.

The draw is that it can run end to end on your own hardware — Ollama for embeddings and reranking, Qdrant in a container — so the repository is never uploaded to index it. The MCP server is a thin surface over the CLI: index once, then the client searches by meaning, reads outlines, or checks how fresh the index is.

What it is

A vector-embedding code search tool that works as a CLI and as an HTTP MCP server. `codebase index --serve` starts the server at `http://localhost:3001/mcp`, and `codebase stdio --server-url=http://localhost:3001/mcp` bridges it for clients that need stdio. Parsing is Tree-sitter across 40+ languages, the vector store is Qdrant, and with Ollama for embedding and reranking the whole pipeline runs offline. It is a fork of Roo Code.

What you get

Semantic search over the indexed repository, returning the matching spans with their line ranges — `search_codebase`
Index statistics, so you can tell whether what the assistant is searching is current — `get_codebase_stats`
Configuration read and changed from the client rather than the shell — `configure_codebase`
Structured outlines of files — functions, classes and methods, optionally with generated summaries per symbol — `outline_codebase`
Call graph analysis from the CLI: a single function's call tree at default depth `3`, connection paths between several functions at default depth `10`, wildcard queries like `--query="get*"`, and Cytoscape.js export with `--viz` or an interactive viewer with `--open`
Call graphs across TypeScript/JavaScript, Python, Java, C/C++, C#, Rust and Go
Search you can narrow: glob path filters with brace expansion, `--limit` up to a maximum of 50, and `--min-score` between 0 and 1
LLM reranking of results on a 0-10 scale, local through Ollama or via an OpenAI-compatible provider — `rerankerEnabled`, `rerankerProvider`, `rerankerMinScore`
Embedding providers to choose from: OpenAI, Ollama, Jina, Gemini, Mistral, OpenRouter and Vercel
File watching that updates the index as the code changes, and a dependency analysis cache the README credits with 10-50x faster re-analysis

Requirements

Install with `npm install -g @autodev/codebase` (1.0.1). Two services have to be up first: Ollama with an embedding model pulled (`brew install ollama ripgrep`, `ollama serve`, `ollama pull nomic-embed-text`) and Qdrant, which the README runs in Docker on ports 6333 and 6334 with `qdrantUrl` pointing at `http://localhost:6333`. ripgrep is installed alongside Ollama in the same step. Keys are only needed if you pick a hosted embedder — `embedderOpenAiApiKey` or `embedderOpenAiCompatibleApiKey` — and the Ollama path needs none. Settings resolve in order: CLI arguments, then `./autodev-config.json`, then `~/.autodev-cache/autodev-config.json`, then built-in defaults; persistent values go through `codebase config --set`. MIT licensed.

Setup effort

One command — npm install -g @autodev/codebase

Autodev Codebase