Free tools Windows power users keep installed
One-click scans. No signup required.
These 12 self-hosted deep research systems are not interchangeable. Some build recursive research trees, some divide work among agents, some look for gaps in evidence, and others focus on cited search answers or personal knowledge. Choose by how you need research to proceed—and verify where inference and search traffic go before treating a deployment as private.
What counts as deep research?
A system that searches the web several times is not necessarily doing adaptive research. The more useful test is what happens after the first results arrive: does the workflow identify unanswered questions or conflicting evidence, investigate again, and produce a report that shows its sources?
As an Amazon Associate I earn from qualifying purchases.
That distinction matters when comparing tools. A polished chat interface may make research convenient without exposing how it branches or decides to continue. Conversely, an agent framework may offer more control but require more setup. The September 25, 2026 comparison behind this selection describes different approaches; it is an editorial comparison, not an independent hands-on evaluation of every current release.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsHow the 12 systems approach research
The table groups each project by the research pattern attributed to it in that comparison. The categories overlap: they describe a system’s central approach, not an exclusive feature or guarantee about every configuration.
| System | Research pattern | Best fit to evaluate |
|---|---|---|
| GPT Researcher | Recursive research tree: branches breadth and depth into additional research tasks. | Explicitly branching web investigations. |
| Unsloth Studio | Evidence-gap loop: uses findings or drafts to identify evidence still needed. | A workflow centered on finding and addressing gaps. |
| Local Deep Research | Evidence-gap loop; supports multiple research strategies. | Local-first research with strategy choice and documented egress controls. |
| STORM / Co-STORM | Perspective-driven research: generates viewpoints and follow-up questions. | Broadening a topic through multiple perspectives. |
| DeerFlow | Planner plus subagents: decomposes tasks and assigns parallel investigations. | Agent orchestration and parallel research; it is also a general agent platform. |
| Onyx | Multi-step knowledge research across public and private sources. | Research integrated with enterprise search and internal repositories. |
| Open Deep Research | Planner plus subagents: decomposes work and assigns parallel investigations. | Agent orchestration, subject to its reported repository status. |
| Open WebUI | Model-directed iterative search: the model can decide when to search, fetch, cross-check, and continue. | A broad self-hosted model interface where research is one capability. |
| Khoj | Knowledge-first research combining personal or workspace knowledge with external research. | Research that uses retained personal knowledge. |
| SurfSense | Knowledge-first research combining personal or workspace knowledge with external research. | Research that uses workspace knowledge. |
| Vane | Search-first answering, prioritizing cited answers over a deep branching graph. | Cited search answers rather than long-running investigations. |
| Deep Research by lukeswade | Evidence-gap loop: uses findings or drafts to identify evidence still needed. | A workflow organized around identifying missing evidence. |
These descriptions do not establish that every project has the same local-model support, document retrieval, interface, deployment requirements, or license terms. Those details can change, so check the relevant project documentation and license before committing to a deployment.
Which deep research system should you choose?
Start with the job you need done, not the number of search steps a product advertises. Use the comparisons below as a shortlist based on the September 2026 article’s characterization, not as a claim that one system wins a controlled test.
For visible branching and broad web investigation
Evaluate GPT Researcher if you want a recursive research workflow that branches tasks across breadth and depth. STORM / Co-STORM takes a different route: its perspective-driven process is intended to expand a topic by generating viewpoints and follow-up questions. Consider which form of breadth fits the question you ask most often.
For evidence gaps and local-first controls
Compare Unsloth Studio, Local Deep Research, and Deep Research by lukeswade if the key question is whether an initial draft or set of findings can surface evidence still needed. Local Deep Research is the clearest candidate in this group when local-first operation and egress controls are priorities; confirm the specific mode and providers you plan to use.
For parallel agent work
DeerFlow and Open Deep Research both fit the planner-plus-subagents pattern. DeerFlow is described as a general agent platform that can be used for research, rather than only as a research application. The September 2026 comparison reported that Open Deep Research had been archived on August 21, 2026. Check its repository status and maintenance before building a new workflow around it.
For enterprise or personal knowledge
Evaluate Onyx when research needs to work alongside enterprise search and internal repositories. For personal or workspace knowledge paired with external research, compare Khoj and SurfSense. In each case, verify connector coverage, access controls, and how repository permissions are applied; the architectural category alone does not establish those implementation details.
For a general model interface or cited answers
Open WebUI is a candidate when you want a broader self-hosted model interface and research is one of several capabilities. Vane is a better fit to evaluate when the desired outcome is a cited answer from search, rather than a long-running investigation with a branching research graph.
What should you check before self-hosting?
“Self-hosted” describes where some components run; it does not prove that model inference, search queries, or other requests stay on your machine or network. Check each route separately.
- Trace model inference. Identify the configured model endpoint and whether it runs locally or sends prompts to a hosted provider. A local-looking endpoint can still relay requests elsewhere.
- Trace web search. Find which search provider receives queries. Local Deep Research’s documentation says its Private only egress mode uses local sources and normally forces local inference, while also describing limitations for its Strict and private modes. It warns that self-hosted SearXNG still forwards search queries to upstream engines.
- Inspect document and connector access. For private knowledge workflows, check which repositories can be queried, which users can access them, and whether their existing permissions carry through.
- Test the evidence trail. Confirm that reports retain sources and that you can inspect what supports a claim, identify conflicting results, and see whether the system revisits unresolved questions.
- Review deployment and maintenance needs. Distinguish a ready-to-run application from a framework or platform you must configure. Check dependencies, update activity, and operational requirements in the current project materials.
- Read the current license. Terms differ across projects and may matter if you modify a network deployment, embed the software commercially, or redistribute it. Verify the license file for the version you intend to use.
How to interpret the published benchmark figures
Local Deep Research’s README reports about 95% on SimpleQA (n=500) and 77% on xbench-DeepSearch (n=100), using Qwen3.6-27B on one RTX 3090. These are figures reported by the LearningCircuit / Local Deep Research project; the year is not stated in the README. The project links them to an announcement and benchmark dataset, but they are not independent comparative results. They describe that particular model and setup, not a universal score for the software or proof that it outperforms the other systems.
For a separate deployment example, NVIDIA’s Deep Researcher Agent illustrates how application hosting and inference hosting can differ. Its repository describes structured planning, concurrent researchers, bounded source-tool batches, and a writer that produces citation-backed reports, with Docker Compose and Helm assets. In the documented default NVIDIA API Catalog profile, inference is hosted and no local GPU is required; the hardware references apply when models are self-hosted. Treat that distinction as a reminder to inspect the exact profile you deploy, not as evidence that every component or configuration is local.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
