Open Source Alternative to OpenAI Deep Research Without the $200 Subscription

When OpenAI rolled out Deep Research as part of the $200/month Pro plan, developers immediately wanted two things: paying only for actual tokens used and connecting their own models. Spending two hundred bucks for a couple of complex research tasks per week doesn't appeal to most people.

A research team from HKUDS extracted the research agent from their AutoAgent framework and packaged it as a standalone repository HKUDS/Auto-Deep-Research. The project runs locally, automatically sets up an isolated Docker environment, and can browse web pages through custom LLM providers.

What's Inside the Project

Auto-Deep-Research performs multi-step information gathering. You pose a complex question, and the agent builds a plan, searches for articles, visits websites, parses pages, and synthesizes findings into a final report.

The architecture draws from several agent system components:

• Isolated execution environment in Docker for safely running browsers and scripts
• LiteLLM integration, enabling the agent to work with OpenAI, Anthropic, Gemini, Mistral, HuggingFace models, and local weights via vLLM
• Support for models with and without native function calling
• Ability to feed local files to the agent for analysis

On the GAIA benchmark, the project shows solid results among open-source agent implementations.

Quick Start and Running in Docker

You'll need Docker and Python 3.10 installed. The utility automatically pulls the appropriate container image for your processor architecture (x86 or ARM), so there's no need to build a Dockerfile manually.

Installing the package in a virtual environment:

conda create -n auto_deep_research python=3.10
conda activate auto_deep_research
git clone https://github.com/HKUDS/Auto-Deep-Research.git
cd Auto-Deep-Research
pip install -e .

Next, create a .env file with your API keys. You don't need to fill in everything—just add the key for the provider you plan to use.

Running with Different Models

By default, the tool targets claude-3-5-sonnet-20241022, but you can override the model right before the launch command via environment variables.

If running with Anthropic:

# в .env задаем ANTHROPIC_API_KEY
auto deep-research

For OpenAI:

# в .env задаем OPENAI_API_KEY
COMPLETION_MODEL=gpt-4o auto deep-research

If you want to run cheap DeepSeek-R1 through OpenRouter or the official API:

# через OpenRouter
OPENROUTER_API_KEY=your_key COMPLETION_MODEL=openrouter/deepseek/deepseek-r1 auto deep-research

# напрямую через DeepSeek
DEEPSEEK_API_KEY=your_key COMPLETION_MODEL=deepseek/deepseek-chat auto deep-research

Groq with fast distillates is also supported:

GROQ_API_KEY=your_key COMPLETION_MODEL=groq/deepseek-r1-distill-llama-70b auto deep-research

The utility accepts --container_name and --port parameters if the default port 12346 is already in use by another service.

Practical Details: Authentication and Files

One common pain point with research agents is paywalls and restricted sites. When the bot hits a page requiring login, it gets stuck.

The authors added a way to pass your browser cookies directly into the container running the agent. The repository has a metachain/environment/cookie_json/ folder where you can drop exported cookies in JSON format. After that, the agent can surf closed portals or scientific databases under your login.

You can also load your own documents. The agent will cross-reference data from the uploaded file with information from the web and factor in the context when formulating conclusions.

Limitations and Rough Edges

The project is fresh, so there are plenty of rough edges:

• No full web interface yet—everything runs through the console (developers promise to add GUI in upcoming releases)
• The repository documentation is sparse; configuration nuances sometimes require digging into the AutoAgent package source code
• If the agent loops on a complex page, token consumption with reasoning models can be surprisingly high

Who Should Try It

Auto-Deep-Research is useful for anyone who regularly needs detailed technical or product reviews but doesn't want to overpay for OpenAI's Pro plan. Pairing it with a local vLLM or an affordable DeepSeek API turns the project into a handy work tool for pennies. If you already have Docker set up, deployment takes literally five minutes.