Teach your LLM the codebase it never trained on

What it does VectorCode embeds and indexes a code repository into a local vector database, then surfaces task-relevant snippets to inject into LLM prompts. It ships as a CLI indexer and a Neovim plugin with Lua APIs for building AI-assisted workflows. The target is codebases the model has never seen—private, niche, or newer than the training cutoff.

The interesting bit The tool chunks code using tree-sitter for semantic and syntax awareness rather than naive text splitting, and it exposes an MCP server so compatible Neovim plugins can query the repo like a tool. That turns “find relevant code” from a manual copy-paste chore into something the LLM can request programmatically.

Key highlights

• Tree-sitter-based chunking with configurable metadata and document selection
• Neovim plugin with Lua APIs, plus integrations for codecompanion.nvim and minuet-ai.nvim
• Model Context Protocol (MCP) support for joint search across collections
• Respects .gitignore and auto-detects project roots via .git or .vectorcode/ anchors
• Local-first storage via ChromaDB’s persistent client

Caveats

• Explicitly beta quality; the author warns of rapid iteration and breaking interface changes before a 1.0.0 release
• Remote ChromaDB authentication and query rewriting are still listed as unfinished TODOs

Verdict Try it if you want your Neovim AI plugins to read your actual repo. Look elsewhere if you need a stable, remote-first, or IDE-agnostic solution right now.