What is local-deep-research?

Local-deep-research, or ldr, is a command line research agent that answers a question by automatically planning, searching the web, reading source pages, and writing a cited report, all without sending your question or data to any cloud AI service. Every AI step runs on a local model through Ollama, llama.cpp, or vLLM, so nothing leaves your machine. When you run ldr ask followed by your question, the tool breaks it into a set of sub-questions in what it calls the Planner stage, then searches the web and fetches relevant pages for each sub-question in parallel during the Researcher stage. It extracts the most relevant passages using BM25 and dense retrieval, two complementary ways to find relevant text in a body of documents, then stitches all the answers into one coherent report during the Synthesizer stage. Finally, a Citer stage verifies every claim has a matching source and flags anything that does not, producing a markdown report with inline numbered citations, an open questions section, and an unsupported claims appendix. Web search runs against SearXNG, a self-hosted, rate limit free search engine recommended for serious use, followed by DuckDuckGo, then the Brave Search API, tried in that order. The tool auto-detects available RAM and picks an appropriately sized model from the qwen2.5 family by default, ranging from a 7 billion parameter model for 8 GB of RAM up to a 32 billion parameter model for 32 to 48 GB. This is useful for people who want deep, sourced research without a subscription or API costs, who care about privacy, or who work in an air-gapped environment with a self-hosted search backend and pre-downloaded models. It is a Python command line project, pre-alpha, released under the Apache 2.0 license.