The archive · Product Ideas · Product decision · 2024–2026
LLocalSearch builds a Perplexity-style answer engine that runs fully on your own machine
A local LLM chooses search tools recursively to answer from the live web — no API keys, no sponsored placement, live logs of every step.
nilsherzig
What it had to solve
Perplexity had made LLM web search mainstream, but the economics of AI answers were turning into media placement — OpenAI's pitch to publishers promised priority placement and richer brand expression inside chat conversations. The developer behind LLocalSearch wanted the same question-answering experience with no API keys and no sponsored results, built locally on ordinary consumer hardware.
How it works
LLocalSearch started from a simple refusal: the developer did not want his answers manipulated by the highest bidder. The README pointed to AI companies selling media houses priority placement and richer link treatment inside chat conversations, and offered a less discriminatory alternative — a Perplexity-style engine where the model, the search and the session all run locally, with no API keys at all.
The mechanism was a wrapper that lets a locally running language model choose from a set of tools. The model could search the web for current information, read what came back, and search again recursively — conditions, loops and multiple tool calls inside one conversation. Because the demo ran on a 7B-class model and a roughly 300€ GPU, the bar was a hobbyist's machine, not a datacenter.
The design also made the agent legible: live logs and links in each answer showed what the model was doing and which pages its claims rested on, with follow-up questions supported in the same session. The Show HN of 3 April 2024 drew 669 points and 133 comments, and the repository grew to 5.9k stars before the owner archived it in June 2026 to rebuild the project in a private beta.
Why it lands
- Ownership was the whole pitch: no API keys, no accounts and no cloud dependency, so the model, the search and the conversation all live on the user's machine.
- Recursion made a small model useful: instead of one prompt, the model could call tools repeatedly and refine its answer against the live web rather than its training cutoff.
- Transparency was the trust move where incumbents sold placement: live logs and cited links let the user audit every step instead of accepting a black-box answer.
- Low-end hardware set the bar: the demo ran on a ~300€ GPU, turning local AI search from a datacenter fantasy into something an individual could run at home.
What it did
The Show HN drew 669 points and 133 comments on 3 April 2024, the strongest debut in its batch, and the repository reached 5.9k stars before the owner archived it on 1 June 2026 while announcing a rewrite in private beta rather than publishing unfinished software.
What you can take
When the incumbent sells placement inside the answer, counter with ownership: run the model, search and audit trail on the user's hardware, and expose the reasoning.
Since then
The project drew immediate attention and forks — 5.9k stars, 363 forks and 182 commits before it was archived — and the launch thread filled with setup questions about hardware, models and quantization, which the author answered in public. On 1 June 2026 the repository was archived as read-only with a note that the version had not been under development for over a year; the author said he was working on a rewrite and relaunch in a private beta, gathering feedback before publishing rather than releasing incomplete software.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card