Inspiration
I kept noticing that LLMs just talk. You ask them something, they answer, done. I wanted to build something where the AI actually does the task — searches something, saves a file, figures out what steps to take. That idea became AgentDesk.
What it does
AgentDesk is an AI workspace where you type a task and the agent figures out what tools to use and executes them. It shows you exactly what it did via action chips in the UI.
Tools it has:
- web_search — searches the web via DuckDuckGo
- read_file — reads a file from the server
- write_file — writes content to a file
- list_files — lists files in a directory
How I built it
React frontend with a dark chat UI, FastAPI backend, and a custom agent loop that runs on LLaMA 3.3-70b via Groq. The agent loop is the core — it checks if the LLM wants to call a tool, runs it, passes the result back, and repeats until there's a final answer. Deployed on Vercel and Render.
Challenges
Getting the agent loop to stop looping was the main headache. The LLM kept calling web_search over and over instead of just answering. Fixed it by tightening the system prompt and adding a hard iteration cap.
CORS between Vercel and Render also took longer than expected to get right.
What I learned
Building a tool-calling agent without LangChain taught me a lot about how these loops actually work under the hood. Also learned how MCP-style tool dispatch works at a fundamental level, and got a lot more comfortable with full-stack deployment.
Log in or sign up for Devpost to join the conversation.