...

How we built Jaqpot's AI agent and MCP server

avatar
Alex Arvanitidis@alarv
17 days ago
· 6 min read

Jaqpot AI Agent and MCP Server

How we built Jaqpot's AI agent and MCP server

At Jaqpot, we want researchers to not only upload and run models but also use them seamlessly inside AI workflows. This led us to two experiments:

  1. An AI agent built with Hugging Face's smolagents.
  2. A Model Context Protocol (MCP) server that plugs Jaqpot into assistants like Claude Desktop.

The AI agent

The AI agent we built (QSARion-smolagents) runs on top of smolagents and the jaqpot-python-sdk.

It defines tools such as:

  • retrieve_smiles_tool → fetches SMILES strings for compounds using PubChem.
  • qsartoolbox_get_property_value_tool → queries Jaqpot's QSAR Toolbox models given a SMILES string and a model name, returning the predicted value, units, and domain of applicability.

The agent is wrapped in a CodeAgent with an OpenAI-compatible model backend, combined with extra tools like DuckDuckGo search and image generation. We connected it to a Gradio UI, so users can chat with the agent, request predictions for compounds, and see results directly.

This setup turned Jaqpot into something more interactive: a conversational interface for exploring chemical properties, predictions, and models.

The MCP server

Next, we took the same SDK foundation and built the jaqpot-mcp-server.

The server implements the MCP standard, exposing Jaqpot as tools any MCP-compatible client can call. Out of the box, it provides:

  • jaqpot_predict → run predictions on a model by ID.
  • jaqpot_search_models → search for models by description or features.
  • jaqpot_get_model_info → fetch metadata for a given model.

Configuration is straightforward: add the server to your claude_desktop_config.json and provide your Jaqpot API key and secret. The server can be run either with uv run server.py or plain python server.py.

The hosting struggle

When we implemented this in April 2025, MCP servers didn't yet support authentication. Since Jaqpot requires an API key and secret, hosting the MCP server online wasn't an option.

Our workaround: run the MCP server locally and pass in the keys via environment variables. This meant every user could securely connect their own Jaqpot account, but it wasn't as smooth as having a shared hosted server.

We expect this to change soon as the MCP community adds authentication support and patterns for safe remote hosting.

Why both matter

Together, these two pieces — the AI agent and the MCP server — make Jaqpot available in new contexts:

  • The AI agent makes Jaqpot predictions accessible through natural-language interaction.
  • The MCP server lets you call Jaqpot directly from IDEs or AI assistants.

They both build on the same SDK, showing how Jaqpot can flexibly fit into research workflows, from conversational exploration to developer automation.

What's next

We'll keep extending both sides:

  • Add more QSAR models and tools to the agent.
  • Add more MCP commands beyond predictions and search.
  • Migrate to hosted setups once secure MCP authentication becomes available.

With these steps, Jaqpot becomes more than a web platform — it becomes a component of the AI ecosystem.

Try it today

Ready to use Jaqpot in your AI workflows?

Get help

Questions? We're here to help: