AgentConn
L

Local Deep Research

Research Free

About Local Deep Research

LearningCircuit/local-deep-research is the privacy-respecting answer to OpenAI Deep Research and Perplexity's research mode. Everything runs locally and encrypted — no telemetry, no third-party API calls unless you explicitly configure cloud LLMs. The headline benchmark is ~95% on SimpleQA tested on Qwen3.6-27B running on a single RTX 3090, putting it inside the band that hosted research products achieve at 100× the cost. The agent searches across 10+ sources — arXiv, PubMed, the open web, and your own private documents — and synthesizes citations into a structured report. Picked up +564 stars on May 7, 2026, riding the broader 'local agents that don't phone home' wave alongside Hermes Agent and the DeepSeek-V4 inference stack. MIT licensed, deployable via Docker / Unraid / direct install. The natural fit is for security researchers, journalists, healthcare professionals, and anyone who needs deep research without their queries leaving the machine.

Key Features

  • 100% local execution — no telemetry, no third-party calls unless you opt in to cloud LLMs
  • ~95% on SimpleQA with Qwen3.6-27B on a single 3090
  • Supports Ollama, llama.cpp, OpenAI, Anthropic, Google, and any OpenAI-compatible endpoint
  • 10+ search engines wired in: arXiv, PubMed, web, and your own indexed private documents
  • Encrypted local storage — research session history stays on disk, encrypted
  • Docker, Unraid, and direct-install deployment paths

Overview

Local Deep Research solves the privacy gap in the deep-research agent category. OpenAI Deep Research, Perplexity, and similar hosted services run your queries through their infrastructure — fine for most users, but a non-starter for journalists, healthcare workers, security researchers, and anyone whose queries themselves contain confidential information.

LearningCircuit’s project ships an end-to-end deep-research loop — query decomposition, multi-source search, evidence synthesis, citation tracking — that runs against any local LLM via Ollama or llama.cpp. The benchmark headline is ~95% on SimpleQA with Qwen3.6-27B on a single RTX 3090, which puts the local stack inside the same accuracy band as the hosted alternatives at orders-of-magnitude lower marginal cost.

The repo accelerated this week alongside the broader local-agents wave: DeepSeek-TUI hitting #1 on GitHub trending, Hermes Agent’s local-AGI stack continuing to pick up stars, and the antirez/ds4 Metal inference engine putting V4 Flash on a MacBook with 128GB of unified memory. Local Deep Research is the research-flavored corner of that wave.

Who It’s For

  • Journalists with sources who don’t want every query logged on a hosted service
  • Healthcare and legal professionals with HIPAA / privilege constraints
  • Security researchers investigating sensitive topics
  • Anyone with a 3090 / 4090 / Apple Silicon Mac who’d rather run inference at electricity cost than pay per-query

Integration

Local Deep Research is a standalone agent — install via Docker, point it at your local Ollama or llama.cpp endpoint, and it handles the rest. It complements rather than replaces coding-agent harnesses; if you’re running DeepSeek-TUI for code work, Local Deep Research slots in alongside it for the research half of your workflow.

Similar Agents