local-deep-research: Free OSS Twin of Deep Research

Deep Research — the agentic research mode that browses, reads, cross-references, and writes you a 3,000-word brief — was the genuinely useful thing OpenAI shipped in 2025. It's also, depending on which tier you're on, somewhere between $20 and $200 per month, with rate limits that get tight once you actually start using it for work.

`learningcircuit/local-deep-research` is the open-source answer. It hit 7.9k stars this month, gaining roughly 3,600 of them in the last 30 days, which is the kind of curve that usually means a real wedge has opened up. The pitch is the obvious one: same loop, your hardware, your API keys, your data.

What's actually in the box

The thing is more general-purpose than the name suggests. It runs in a few modes — Quick Summary for the "give me the gist of this topic" use case, Detailed Research for the multi-hour agentic loop, and Report Generation for the structured-output version.

Search-engine coverage is the part that surprised me:

• Academic: arXiv, PubMed, Semantic Scholar, NASA ADS
• General web: Wikipedia, SearXNG, Wayback Machine
• Technical: GitHub code search, Elasticsearch
• Premium when you have keys: Tavily, Brave, Google via SerpAPI
• Custom: local PDFs, LangChain retrievers, your own corpus

That last bucket is the one that makes this interesting beyond "cheap Deep Research clone." You can point it at a folder of internal docs, a Notion export, a competitor's site map, and have it run the same research loop over private data — something the hosted Deep Research subscriptions explicitly can't do.

LLM flexibility is where the cost story lives

It runs against anything that speaks the OpenAI API shape — Ollama, LM Studio, llama.cpp on the local side; OpenAI, Anthropic, Gemini, OpenRouter on the hosted side. The project benchmarks ~95% on SimpleQA with a Qwen3.6-27B running on a single 3090, which is a real result, not vibes.

Where it's genuinely better than the hosted thing

Two real advantages, neither of which is "it's free."

First, the search engine list. Hosted Deep Research products use whatever index their parent company licenses. This thing lets you mix arXiv and PubMed with your own SearXNG instance and a local Elasticsearch over your internal data, in one query. That's the kind of stack a research-heavy SMB actually wants — a market that the hosted products are not really serving.

Second, encrypted per-user databases with SQLCipher and AES-256. Boring detail, but it's the difference between "I can run this for a client engagement" and "I need to read our DPA before I touch this."

Where it isn't

Setup is real work. Docker is the cleanest path, but if you want local LLMs doing actual heavy lifting you're in Ollama-config land, and you need a GPU that can hold a 27B-parameter model. The hosted product's whole pitch is that you click a button and it works.

Output quality on local models is also lower than GPT-5.5 or Claude Opus 4.7, full stop. The 95% SimpleQA benchmark is on factual lookup; for synthesis-heavy briefs the gap widens. If you're willing to point it at a hosted API, you close most of that gap and still beat the subscription price.

"Hosted Deep Research is the right choice for occasional users. For anyone running this loop more than ten times a month, the per-query math eats you alive — and this is the OSS escape hatch."

v1.6.12 shipped this week — 156 releases in, which is a healthier release cadence than half the dev-tool startups I check on. For an indie hacker or small dev shop doing structured research as part of the actual work — competitive briefs, vertical landing-page research, technical scoping — this is the cheap version of a tool you'd otherwise be paying for monthly. Worth a weekend.