Auto-Deep-Research is an open-source personal AI assistant positioned as an alternative to OpenAI’s Deep Research, released by the research team of Chao Huang at the HKU Data Science Institute (HKU IDS), University of Hong Kong. It was published on February 16, 2025 as the first practical application built on the team’s agent framework, AutoAgent. It combines web search, browser automation, and file parsing to carry out research tasks autonomously. Instead of a fixed monthly subscription, it runs on pay-as-you-go usage of an LLM API key you supply yourself, and you launch it by typing auto deep-research in your terminal.
Key Features
- Pay-as-you-go with your own API key: Through LiteLLM, it supports OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, Hugging Face, OpenRouter, and even OpenAI-compatible endpoints such as xAI’s Grok. You switch models by setting the
COMPLETION_MODELenvironment variable (the default is Claude 3.5 Sonnet) - Launches with a single command: When you run it, a container image matching your machine’s architecture is fetched automatically and an isolated environment for the agent spins up. No manual image pull, no configuration file to write
- Works with models that lack function calling: It handles both models with and without function-calling support, and which mode to use is inferred automatically from the model name. That puts inexpensive small models and local models without function calling on the table as well
- Reads a wide range of file formats: Its dependencies include parsers for PDF, Word, and PowerPoint, plus speech recognition (faster-whisper), video processing (MoviePy), and YouTube transcript retrieval, so non-text material can also be part of your research
- Browses with a real browser: It ships with a browser environment powered by BrowserGym and Playwright, so it can reach pages a search API alone cannot. Import your cookies and it can handle sites that require a login
How to Think About Cost
There is no paid plan. Here is where the money goes.
- The software itself: Free. Open source, published on GitHub
- Actual expense: Only the API usage fees of the LLM provider you have signed up with. The cost per run scales with the model’s unit price and the depth of the research (number of exploration steps and volume of pages read)
- Runtime environment: It runs on Docker on your own machine, so there are no server costs
At the time of release, OpenAI’s Deep Research was available only on the $200-per-month ChatGPT Pro plan, and this project was announced as an alternative that “does the same thing without an expensive fixed subscription.” Deep Research has since rolled out to lower-tier plans, so which is cheaper now depends on how often you run research and which models you pick. A few times a month, pay-as-you-go wins; running it daily, a flat fee is easier to budget.
Pros and Cons
✅ Pros
- No fixed monthly cost; you pay only for the API usage you consume
- Free choice of model. You can do preliminary research on a cheap model and reserve a high-end model for the investigations that matter
- It runs in a local Docker environment, so the files you research and any intermediate data never leave for an external SaaS
- The code is public, so you can inspect and modify the agent’s behavior and prompts yourself
- The underlying AutoAgent has a preprint (arXiv:2502.05957), so you can trace the design rationale back to the source. The paper reports outperforming existing approaches on the GAIA benchmark, which measures general assistant capability, and the lab’s announcement page states that it “maintains a top position on GAIA with performance comparable to Deep Research” (neither cites specific scores)
⚠️ Cons
- CLI only, no GUI. The README says a web interface is “under development,” but none is available at this time
- It assumes a Python 3.10+ environment and Docker, which is a high bar for non-engineers
- Both the project page and the documentation site linked from the README return 404 (as of August 2026). In practice, the README is the entire setup guide
- The package configuration (setup.cfg) lists MIT as the license and the underlying AutoAgent is also MIT, but there is no LICENSE file at the repository root. Worth clarifying the rights situation before commercial use
- As of this writing (August 2026), the latest commit dates to October 2025. The pace of development has slowed since the initial release
Comparison with Similar Services
| Criteria | Auto-Deep-Research | OpenAI Deep Research | Perplexity Deep Research | GPT Researcher |
|---|---|---|---|---|
| Form | OSS (CLI + Docker) | ChatGPT feature | Perplexity feature | OSS (CLI + Web UI) |
| Where it runs | Your own machine | Cloud | Cloud | Your own machine / your own server |
| Billing | LLM API usage only | Subscription | Subscription | LLM API usage only |
| Model choice | Free choice among major providers | Fixed to OpenAI models | Chosen by the provider | Free choice among major providers |
| License | MIT stated in setup.cfg (no LICENSE file) | ─ | ─ | Apache-2.0 |
| Ease of setup | Requires Docker + Python setup | Account only | Account only | Requires Python setup |
Who Is It For
- People who already hold an LLM API key and do not want to add another monthly subscription for research
- Developers who want to choose their own research model and switch between cheap and high-performance models
- People who want to read the agent’s internals and adapt it to their own workflow
- People who want to keep the material they research inside their local environment
- People who want a working agent application as a starting point for exploring AutoAgent
Summary
Auto-Deep-Research is an open-source implementation that lets you run an autonomous research agent from your own environment, on a model of your choosing. Having a paper and a framework behind it is reassuring. On the other hand, the barrier to entry is not low: CLI only, Docker required, and an official documentation site that does not work. If you are comfortable in a terminal and hold an API key, it is worth a try. Start with a small research task on an inexpensive model.