Skip to main content
PromptQuorum
Home/Power Local LLM/Local Deep Research Review 2026: Self-Hosted, Cited AI Research Agent
Local AI Agents & Tool Use

Local Deep Research Review 2026: Self-Hosted, Cited AI Research Agent

Β·12 min readΒ·By Hans Kuepper Β· Founder of PromptQuorum, multi-model AI dispatch tool Β· PromptQuorum

Local Deep Research is a free, open-source, self-hosted AI research agent (source at github.com/LearningCircuit/local-deep-research) that runs multi-step research across sources like arXiv, PubMed, Wikipedia, GitHub, and the general web, then produces a cited report using an LLM you configure yourself. It is licensed under MIT and has no paid tier β€” the GitHub repository has no pricing page, and the only costs are whatever LLM provider or search API you choose to connect. It installs via pip install local-deep-research, works with local models through Ollama, LM Studio, or llama.cpp, and also supports cloud providers (OpenAI, Anthropic, Google Gemini, OpenRouter) and premium search APIs (Tavily, SerpAPI, Brave Search) as opt-in additions.

Local Deep Research (github.com/LearningCircuit/local-deep-research) is a free, open-source Python tool that runs multi-step, agentic research and produces cited reports by querying sources like arXiv, PubMed, Wikipedia, and the web, then summarizing findings using an LLM you configure β€” a local model via Ollama or llama.cpp, or a cloud API. It installs via pip install local-deep-research and exposes a web UI, a Python library, a CLI, an HTTP API, and an MCP server for Claude Desktop and Claude Code integration. This review covers what it actually does, how to install and configure it, and how it compares to other self-hosted research and search agents.

Key Takeaways

  • Local Deep Research is free and open-source; the official LICENSE file is MIT
  • Runs multi-step, agentic research (not a single search-and-summarize call) and produces reports with citations back to sources
  • Works with local LLMs β€” Ollama, LM Studio, and llama.cpp are supported backends β€” or cloud providers (OpenAI, Anthropic Claude, Google Gemini, OpenRouter, Requesty)
  • Free search sources: arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, the Wayback Machine, and The Guardian; premium options (Tavily, Google via SerpAPI, Brave Search) are opt-in and require their own API keys
  • Interfaces: a web UI (localhost:5000), a Python library (quick_query(), LDRClient), a CLI, an authenticated REST API, and an MCP server for Claude Desktop and Claude Code
  • Per-user, AES-256 encrypted databases, with no telemetry, analytics, or tracking built into the tool itself, per the project's own documentation
  • Available via pip install local-deep-research or Docker Compose (which also runs Ollama and SearXNG alongside it)
  • GitHub repository (github.com/LearningCircuit/local-deep-research) shows roughly 9,100 stars as of September 2026

πŸ“ In One Sentence

Local Deep Research is a free, open-source, self-hosted AI research agent that runs multi-step research across academic and web sources and produces a cited report using an LLM you configure, local or cloud.

πŸ’¬ In Plain Terms

Instead of asking a chatbot one question and getting one unverified answer, Local Deep Research breaks a research question into multiple steps, searches several sources (like arXiv and PubMed for academic topics, or the open web for general ones), and writes up a report that cites where each claim came from β€” you can run the whole thing on your own machine with a local model like Ollama, so the research question itself never has to leave your computer.

πŸ“ŒNote: This review is the deep-dive companion to Local Deep Research's entry in the Local LLM Software Directory β€” see that page for how it compares at a glance to dozens of other local and self-hosted AI tools.

What Is Local Deep Research?

Local Deep Research is a Python tool for running multi-step, cited AI research, using an LLM and search sources you configure yourself. Its own GitHub description positions it as an "AI-powered research assistant" for deep, agentic investigations, emphasizing that it runs "everything local and encrypted" when configured with local models β€” a design choice aimed at researchers who want a documented, source-cited answer rather than a single unverified chat response.

  • Core function: an agent that plans and executes multi-step research (not a single query), pulling from configurable academic and web sources, then writing a cited report
  • LLM backends: local models via Ollama (default endpoint http://localhost:11434), LM Studio, or llama.cpp; cloud providers OpenAI, Anthropic Claude, Google Gemini, OpenRouter, and Requesty are also supported
  • Search sources: free options include arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, the Wayback Machine, and The Guardian; premium options (Tavily, Google via SerpAPI, Brave Search) require the user's own API keys
  • License: MIT, a permissive open-source license
  • Developer/organization: the LearningCircuit GitHub organization β€” the canonical, actively maintained home of the project's source code, releases, and issue tracker
  • Interfaces: a browser-based web UI, a Python API (quick_query() and LDRClient classes), a CLI for benchmarking and automation, an authenticated REST API, and an MCP server for Claude Desktop and Claude Code

Local Deep Research's Project History and Version Milestones

Local Deep Research has shipped continuously, with the GitHub repository showing thousands of commits on its main branch as of this review, adding search-source integrations, an MCP server, and encrypted per-user databases along the way. The milestones below are drawn from the project's public repository activity β€” check the GitHub repository directly for the complete, current history.

  1. 1
    Core multi-step research engine and CLI
    Why it matters: The foundational agentic research loop β€” planning a query, searching multiple sources, and synthesizing a cited answer β€” rather than a single search-and-summarize call.
  2. 2
    Free academic and web search source integrations (arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, Wayback Machine, The Guardian)
    Why it matters: Gives the agent access to citable academic literature and general web sources without requiring a paid search API.
  3. 3
    Local LLM backend support (Ollama, LM Studio, llama.cpp)
    Why it matters: Lets the entire research loop β€” planning, searching, and report generation β€” run against a model on your own hardware, with no research query sent to a third party.
  4. 4
    Docker Compose deployment with bundled Ollama and SearXNG
    Why it matters: Simplifies self-hosting by bringing up the local model server and a self-hosted search engine (SearXNG) alongside the research agent in one `docker compose up -d` command.
  5. 5
    Per-user AES-256 encrypted databases
    Why it matters: Adds encryption at the per-user database level for research history and results, relevant for anyone self-hosting Local Deep Research for multiple users.
  6. 6
    MCP server for Claude Desktop and Claude Code
    Why it matters: Lets Local Deep Research act as a tool inside Claude Desktop or Claude Code via the Model Context Protocol, rather than only as a standalone web UI or CLI.
  7. 7
    v1.10.3 β€” SearXNG instance URL pinning requirement
    Why it matters: As of this version, using SearXNG's localhost search engines requires pinning the instance URL in configuration β€” documented in the project's `docs/SearXNG-Setup.md`, the most recent version-specific change this review could verify.

What Can You Do With Local Deep Research?

Local Deep Research's feature set centers on turning a research question into a multi-step, cited investigation, with local-model support as a first-class option rather than an afterthought. Here is what each part actually does, per the project's own GitHub README.

  • Research modes β€” quick summaries for a fast overview, detailed research for a deeper multi-source investigation, full report generation, and document analysis over files you provide
  • Multi-step agentic search β€” the agent plans a research approach, queries multiple configured sources, and synthesizes findings into a report with citations back to each source, rather than returning one unverified chat answer
  • Local LLM support β€” run the entire research and writing pipeline against a local model through Ollama (default http://localhost:11434), LM Studio (http://localhost:1234/v1), or llama.cpp (http://localhost:8080/v1), so the research question and results never leave your machine
  • Cloud LLM support β€” connect OpenAI, Anthropic Claude, Google Gemini, OpenRouter (100+ models), or Requesty (100+ models) as an opt-in alternative to a local model
  • Academic and web search sources β€” free sources include arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG (self-hosted meta-search), GitHub, the Wayback Machine, and The Guardian; premium sources (Tavily, Google via SerpAPI, Brave Search) require the user's own API key
  • Custom and local document search β€” search your own local documents, connect LangChain retrievers, or combine multiple sources in a meta-search configuration
  • News subscriptions β€” set up automated, recurring research digests on a topic with a customizable delivery schedule
  • REST API, analytics dashboard, and benchmarking tools β€” an authenticated HTTP API with CSRF protection, a dashboard for tracking research runs, and built-in benchmarking utilities for comparing model/search configurations
  • Export and integration β€” export research reports to PDF or Markdown, and connect Local Deep Research to Claude Desktop or Claude Code as an MCP server

Usage Examples: Three Ways to Use Local Deep Research

These are concrete workflows built from Local Deep Research's documented features above β€” not hypothetical use cases.

Local Deep Research Pricing: Is It Really Free?

Yes β€” Local Deep Research itself has no paid tier. The GitHub repository has no pricing page, and its MIT license applies to the entire codebase with no feature gated behind payment. The costs you might actually incur come from the LLM provider and search sources you choose to connect, not from Local Deep Research itself.

Local Deep Research software

Cost:
Free, MIT licensed, no usage limits imposed by the tool itself

Local LLM backend (Ollama, LM Studio, llama.cpp)

Cost:
Free β€” runs on your own hardware, no per-query API cost

Cloud LLM backend (OpenAI, Anthropic, Gemini, OpenRouter, Requesty)

Cost:
Pay the provider directly for API usage; Local Deep Research does not add a markup

Free search sources (arXiv, PubMed, Wikipedia, SearXNG, GitHub, etc.)

Cost:
Free, no API key required for most of these

A fully local setup β€” a local model via Ollama plus free search sources like arXiv, PubMed, and SearXNG β€” can run Local Deep Research at effectively zero ongoing cost beyond your own hardware and electricity. Adding cloud LLM or premium search providers introduces per-usage costs from those providers, entirely separate from Local Deep Research's own MIT-licensed, free codebase.

Local Deep Research vs. Vane

Local Deep Research and Vane (formerly Perplexica) are both self-hosted, open-source alternatives to closed AI search products, but they target different research depths. Vane is built to answer a question quickly with a Perplexity-style cited response; Local Deep Research is built for deeper, multi-step investigations that can take significantly longer but cover more ground.

Primary use case

Local Deep Research:
Multi-step, deep research reports with citations
Vane:
Fast, Perplexity-style answering with citations

Academic sources

Local Deep Research:
arXiv, PubMed, Semantic Scholar built in
Vane:
See the dedicated Vane review for its source list

Local LLM support

Local Deep Research:
Ollama, LM Studio, llama.cpp
Vane:
See the dedicated Vane review for its supported backends

Interfaces

Local Deep Research:
Web UI, Python library, CLI, REST API, MCP server
Vane:
See the dedicated Vane review for its interface options

License

Local Deep Research:
MIT
Vane:
See the dedicated Vane review for current license details

Report depth

Local Deep Research:
Multi-step reports, research modes from quick summary to full report
Vane:
Single-turn or conversational answers, not a dedicated multi-step report mode

If your main need is a quick, cited answer to a specific question while browsing β€” a self-hosted Perplexity replacement β€” Vane fits that use case directly; see the full Vane review. If your main need is a longer, multi-source research report you can export and cite, Local Deep Research's deeper, multi-step design fits better.

Who Should Use Local Deep Research?

Whether Local Deep Research fits depends on whether you need genuinely multi-step, cited research output and are comfortable running a Python tool with your own configured LLM backend.

Local Deep Research vs. Other Research and Agent Tools

Local Deep Research sits in a broader category of self-hosted tools for AI-assisted search and autonomous research/agent workflows. Here is how it compares to other options in that space β€” see the Local LLM Software Directory for the full catalog, and the dedicated Local Deep Research vs. Vane comparison above for the closest head-to-head.

  • Vane (formerly Perplexica) β€” a self-hosted, Perplexity-style AI answering engine built for fast, single-turn cited search rather than multi-step deep reports; see the dedicated comparison section above and the Vane review for full details.
  • Onyx (formerly Danswer) β€” a self-hosted enterprise search and RAG platform built for querying an organization's own internal documents and connectors, a different use case from Local Deep Research's academic/web research focus; see the Onyx review.
  • Agent Zero β€” a general-purpose autonomous agent framework capable of research among many other task types, broader in scope but not specialized in cited academic research the way Local Deep Research is; see the Agent Zero review.
  • BabyAGI β€” an early, influential autonomous task-management agent framework; useful context for how agentic research loops evolved, though less specialized for cited research output than Local Deep Research; see the BabyAGI review.

Common Mistakes When Evaluating Local Deep Research

Most confusion about Local Deep Research comes from assuming it works like a single-question chatbot, misunderstanding which parts of its pipeline actually run locally, or overestimating its reported benchmark numbers as universal.

Frequently Asked Questions

What is Local Deep Research?

Local Deep Research (source at github.com/LearningCircuit/local-deep-research) is a free, open-source Python tool that runs multi-step, agentic research across academic and web sources and produces a cited report using an LLM you configure, local or cloud.

Is Local Deep Research free?

Yes. The tool itself is free and MIT licensed, with no pricing page and no paid tier. If you connect a cloud LLM provider or a premium search API, you pay that provider directly β€” Local Deep Research does not add a markup.

Is Local Deep Research open source? What license does it use?

Yes, it is open source under the MIT license, confirmed via the LICENSE file in its GitHub repository.

Can Local Deep Research run fully offline / fully local?

Yes, if configured that way. Using a local LLM backend (Ollama, LM Studio, or llama.cpp) plus free sources like arXiv, PubMed, Wikipedia, and a self-hosted SearXNG instance keeps the research and generation pipeline on your own hardware. Public search requests to sources like arXiv or Wikipedia still leave your machine to reach those public services, which is a separate consideration from whether your LLM inference is local.

How do I install Local Deep Research?

Install via pip install local-deep-research and run python -m local_deep_research.web.app, or use docker compose up -d with the repository's Docker Compose file, which also brings up Ollama and SearXNG alongside the app.

What LLM providers does Local Deep Research support?

Local backends: Ollama, LM Studio, and llama.cpp. Cloud providers: OpenAI, Anthropic Claude, Google Gemini, OpenRouter (100+ models), and Requesty (100+ models), plus custom OpenAI-compatible and Anthropic-compatible endpoints.

What search sources does Local Deep Research use?

Free sources include arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, the Wayback Machine, and The Guardian. Premium, opt-in sources include Tavily, Google (via SerpAPI), and Brave Search, each requiring the user's own API key.

Does Local Deep Research have an API or MCP support?

Yes. It offers an authenticated REST API with CSRF protection, a Python library (quick_query(), LDRClient), and an MCP server for integration with Claude Desktop and Claude Code.

How is Local Deep Research different from Vane (formerly Perplexica)?

Vane is built for fast, single-turn, Perplexity-style cited answers. Local Deep Research is built for deeper, multi-step research with dedicated research modes (quick summary through full report) and built-in academic sources like arXiv and PubMed β€” see the dedicated comparison section above.

Who develops Local Deep Research?

Local Deep Research is maintained under the LearningCircuit GitHub organization, which hosts the project's active source code, releases, and issue tracker.

Sources

← Back to Power Local LLM