1. Download the code

   git clone https://github.com/IlyaGusev/holosophos.git
   cd holosophos
   

2. Create a .env file at the project root. Typical entries (keys below are dummy):

   # LLM provider keys (use what you need)
   OPENROUTER_API_KEY=sk-or-v1-b8caassdasdasdasdasd978063a
   
   # Tooling keys
   TAVILY_API_KEY=tvly-dev-oasdasdasdasdasdkb5F
   VAST_AI_KEY=4b3448aaaasdasdasdasdas359c31a4c144b
   
   # Optional observability
   PHOENIX_URL=http://localhost:6006
   PHOENIX_PROJECT_NAME=holosophos
   
   # MCP endpoints when running locally
   ACADEMIA_MCP_URL=http://0.0.0.0:5056/mcp
   MLE_KIT_MCP_URL=http://0.0.0.0:5057/mcp
   
   # App options
   MODEL_NAME=openai/gpt-5
   PORT=5055
   

   You will need an actual OPENROUTER_API_KEY, which you will get after registration.

   [synced_block is not supported]

   `TAVILY_API_KEY` should be obtained for free by registration on [https://www.tavily.com](https://www.tavily.com/)
   

   VAST_AI_KEY is optional if you would like to use an external server with GPU for computations (recommended).

3. Create a folder workdir in the main folder

4. Run inside the main folder

   docker compose up --build
   

   If you run this locally on your machine, you will be able to go to http://localhost:5055 for the application (unless a different port is specified)

5. Run this in the main folder

   uv run python -m codearkt.terminal
   

6. Write your prompt for research here, then press Esc, then Enter. For example:

   Draft a 4-page research paper on the connection between ODE splitting schemes and optimization algorithms with:
   - Abstract, Introduction, Methods, Results, Discussion, Conclusion
   - ~1200–1600 words (≈4 pages), clear sectioning
   - 2 figures (placeholders OK) and 1 table
   - ~10 references (APA/ACM style)
   Use LaTeX and compile to PDF. Save all sources and the PDF under /workdir/paper.
   
   A short description is here:
   
   Different ODE splitting schemes induce distinct optimization dynamics; deriving optimizers from higher-order and asymmetric splits can give us new optimization algorithms
   
   in the paper
   
   https://arxiv.org/pdf/2004.08981 
   
   
   The authors present a different view on stochastic optimization, which goes back to the
   splitting schemes for approximate solutions of ODE. In this work, the authors provide a
   connection between stochastic gradient descent approach and first-order splitting
   scheme for ODE. The authors consider the special case of splitting, which is inspired by
   machine learning applications and derive a new upper bound on the global splitting
   error for it. The authors present, that the Kaczmarz method is the limit case of the splitting
   scheme for the unit batch SGD for the linear least squares problem. The authors support
   our findings with systematic empirical studies, which demonstrates, that a more
   accurate solution of local problems leads to the stepsize robustness and provides
   better convergence in time and iterations on the softmax regression problem.
   
   let us study another splitting schemes and derive different optimziation algorithms based on them?
   
   Carefully track experiment time and perform only those experiments you can do, starting from vert small scales
   
   ## Experimental rules:
   Ensure, that measuring starts before running different methods, to ensure, that they started from the same initialization point.
   Fix seeds, consider small linear regression, logistic regression, softmax regression problems, ideally focus on real world dataset, rather than synthetic one
   

Settings: holosophos/settings.py. You can adjust: MODEL_NAME, MAX_COMPLETION_TOKENS, max_iterations for each agent, etc.

Instructions are located at: holosophos/prompts/*.yaml

How to use your own latex template? Put your .sty and a base .tex skeleton into workdir/paper/ . By default, it uses agents4science latex template, but you may prompt it to use your custom template.