Common questions
FAQ
Deployment, training, and integrating MolClass into your own workflows.
How can I train a model?
Not on this server — this instance is configured to serve predictions and search against already-approved models only. Uploading new data and training new models are both disabled here.
To train your own models, download the MolClass code and run it yourself (see the next question).
Where can I download MolClass?
MolClass is on GitHub:
What’s the best way to deploy it?
Docker Compose. The repository ships a complete stack — database, API, SDF import worker, model-training worker, prediction service, and frontend — defined in a single docker-compose.yml.
git clone https://github.com/jwildenhain/molclass.git cd molclass # Create an untracked .env with two required secrets cat > .env <<'EOF' MOLCLASS_DB_ROOT_PASSWORD=<a long random secret> MOLCLASS_DB_PASSWORD=<a different long random secret> EOF docker compose build docker compose up -d
The frontend listens on 127.0.0.1:3000 by default (override with MOLCLASS_FRONTEND_PORT in .env); put a reverse proxy in front of it for a public deployment. See docker-compose.yml for the full set of tunables — per-service memory limits, ports, and worker thread counts.
How can I use it with my own workflows?
MolClass ships an MCP (Model Context Protocol) server, so AI assistants and agent workflows (Claude Desktop, Claude Code, or any MCP-compatible client) can look up compounds and run predictions directly against a MolClass deployment.
It’s in the repository under mcp-server/, with setup instructions in its README. To keep any one deployment from being overwhelmed by automated requests, each MCP server instance self-limits to 100 predictions per day.