localgpu

Run a local Ollama instance with nomic-embed-text, index the repository into an on-disk vector store, and call the search_code MCP tool with a natural-language query. It returns ranked file:line hits with excerpts of at most three lines, all over 127.0.0.1.

Use the localgpu shell command, which starts a proxy translating Anthropic Messages API calls into Ollama's /api/chat and launches a separate claude process with ANTHROPIC_BASE_URL pointed at it. The current session's model is never changed.

Check index_status first: the root may not be indexed, the index may be stale relative to recent edits, or an ignore glob like dist/ or build/ may exclude the file. Also note embedding search finds ideas, so exact symbol names are better handled with Grep.

The index is invalidated because vectors from different embedding models are not comparable. The manifest records the embed model name and every search_code or index_refresh call raises an error on mismatch until you rebuild with a full index refresh.

Two models are resident at once, fighting for VRAM alongside the desktop. Set OLLAMA_MAX_LOADED_MODELS=1 and a short keep_alive on embedding requests so Ollama evicts one model before loading the other; /api/ps confirms what is loaded.

Yes, with a caveat: the shipped qwen2.5-coder model writes tool calls as JSON text, and the proxy promotes them to real tool_use blocks only for tools the request offered. Images, thinking blocks, and prompt caching are not carried across the translation.