Agent4Molecule

Frontend initial setup

1. Clone the repository to the system/server. Our existing MCP server submit jobs to the PSC server.
2. Install the required packages to run the agent and frontend code: pip install -r requirements.txt
3. Modify line 21 and update USER_DIR to a directory where all agent states and chat histories can be saved.
4. In src/mcp_agent/, modify all paths at the top of enzygen_server.py,heme_binder_server.py, and ppdiff_server.py to point to the right package locations on your system.
5. In src/mcp_agent/utils, modify all paths in each .py file to be valid on your system.
6. Make sure you have a valid gemini API key before running the UI.
7. Follow the instructions to set up the environment for enzygen and heme binder.

Running the Frontend

1. Modify line 121 of src/chat_interface.py with the path of the MCP server you wish to run the agent with. Currently, it is set up with enzygen_server.py.
2. Before running the frontend, export GEMINI_API_KEY with your key. Do not hardcode and push this token to github.
3. To run the frontend, run the following. The port can be changed from 8050 as needed.

streamlit run src/chat_interface.py --server.port=8050 --server.address=0.0.0.0
streamlit run src/chat_interface.py --server.port=<PORT> --server.address=0.0.0.0

The UI can now be accessed at http://localhost:8050/, or http://localhost:/.

4. If the UI is launched from a compute node instead of a login node, perform the following steps: a. Request the Interact GPU and take note of the compute node number. On the PSC system, this can range from v001 to v024. b. Export GEMINI_API_KEY with your key on the compute node's environment. c. Run the following command with a port such as 8050: streamlit run src/chat_interface.py --server.port=<GPU_PORT> --server.address=0.0.0.0 d. On your local terminal, set up another connection to the remote server using port forwarding with a new local port. For example, if using PSC, a compute node of v007, and a local port of 9001, the command would be

   ssh -L 9001:v007:8050 username@bridges2.psc.edu
   ssh -L LOCAL_PORT:GPU_NO:GPU_PORT username@server_address
   

   e. The UI is now accessed at http://localhost:9001/, or http://localhost:<LOCAL_PORT>/.

Adding custom MCP servers to the UI

To make an MCP server work our UI setup, we recommend adding descriptions for each function and parameter. Specifically, each non-internal function (those decorated with @mcp.tool()) must return a dictionary with the following key:value pairs:

• "status":
  • "successful": The function successfully finished.
  • "error": There was some error in the code execution. Use this flag when an exception is caught in a try-catch block.
• "answer":
  • Verbose answer of what output was produced. Include the file path and any other pertinent information. This field is ultimately displayed on the UI side.
• "visualize":
  • "none": No visualization expected for this function
  • "plot": The frontend uses st.image() to render the file in the "file_path" key.
  • "molecule" : The frontend uses molviewspec to render the file in the "file_path" key.
  • "docking": The frontend adds a ligand to the molecular visualization.
• "message_render":
  • "text": Use regular st.markdown() to display the text in "answer".
  • "table": If the "answer" key is in some columnar format, use st.table() to display the result.
• "file_path":
  • A single or list of files to be visualized. Use "none" if there is nothing to visualize.

For examples, refer to the existing MCP servers in mcp_agent (enzygen_server.py, heme_binder_server.py, ppdiff_server.py).