A local search agent that treats OpenAI's pitch deck as a threat model

What it does

LLocalSearch is a Go-based wrapper around local LLMs that gives them a toolbox for web search. You ask a question, and the model recursively decides which tools to call, pulling live information from the internet until it cobbles together an answer. The whole thing stays on your hardware—no OpenAI, no Google, no API keys.

The interesting bit

The project wears its politics on its sleeve. The author explicitly cites OpenAI’s “preferred publisher” pitch—where media partners buy priority placement in chat answers—as the motivation for building a search pipeline no advertiser can bid on. That makes the recursive agent loop feel less like an AI demo and more like an ad-blocker for LLM search.

Key highlights

• Runs entirely locally on “low-end” hardware; the demo was recorded on a 300€ GPU.
• Recursive tool use: the LLM can chain multiple web searches and follow-up questions before answering.
• Live logs and source links are surfaced in the UI, so you can audit what the agent read instead of trusting a black-box summary.
• Supports follow-up questions and has a mobile-friendly interface with dark/light mode.
• Written in Go, using the langchaingo library to orchestrate the agent loop.

Caveats

• The repository has been stale for over a year; the author is now rewriting it in private beta rather than updating this codebase.
• Llama 3 support is buggy because langchaingo ignores stop words, causing the model to hallucinate at the end of turns; a patch exists on an experiments branch but is waiting on upstream.
• Chat history and long-term memory are still on the roadmap, requiring internal data-structure refactors that have not landed.

Verdict

Worth a look if you want a transparent, local alternative to commercial AI search and don’t mind running slightly dusty Go code. If you need production reliability or modern model support today, wait for the promised rewrite or bring your own patches.