🤖 Tool-Calling Agent With Local LLM

This interactive web application allows you to experiment with AI agents in your browser, using OpenAI Agents Python SDK and Pyodide. You can customize agent behavior, test different prompts, and see responses in real-time. Check the mozilla-ai/wasm-agents-blueprint GitHub repository for more information.

How it works: this application runs a local agent which can make use of the following tools:

• count_character_occurrences which counts the occurrences of a given character inside a word
• visit_webpage which visits a webpage at the provided url and reads its content as a markdown string
• search_tavily which performs web searches using Tavily API (requires TAVILY_API_KEY in config.js)

While the former tool is quite trivial and is mainly used to show how to address the "r in strawberry" issue, the latter two provide the LLM with the capability of accessing up-to-date information on the Web.

• Configure: Make sure the Local LLM Server Configuration parameters are ok for your setup. In particular, the default expects you to have Ollama running on your system with the qwen3:8b model installed. You can also click the LM Studio preset button if you are using LM Studio, and make sure to update your model name accordingly. Optionally, add your TAVILY_API_KEY to config.js to enable web search functionality.
  NOTE: if you are using LM Studio with a thinking model and are getting tool calls directly in the model's response, disable thinking in the "Edit model default parameters" section.
• Initialize: Set up the Python environment with Pyodide and the OpenAI agents framework by clicking on the Initialize Pyodide Environment button
• Customize: Choose one of the suggested prompts or create new ones in the text fields below. (hint: you can also explicitly set/unset qwen3's "think mode" by prepending /think or /no_think to the prompt).
• Run: Click on the Run Agent button to send your prompt to the agent and see what happens