StressBall

Expanded Elevator
Architecture

Inspiration

When's the last time you had an excruciatingly long phone call with someone who you didn't really want to talk to? Whether it is your cellphone carrier, healthcare provider, or customer support, we've all had our fair share of annoying phone calls. We think that this should be easier: as easy as sending a text.

What it does

StressBall is a voice agent that calls businesses on your behalf to do what you would otherwise do yourself. It's as easy as texting Poke or your favorite MCP-enabled tool, and StressBall will:

Reclaim time in phone trees (it navigates them)
Save energy (it never gets tired)
Reduce anxiety (it stays calm the entire time) while making a phone call, so you can do more important things instead!

How we built it

StressBall is a serverless Model Context Protocol (MCP) server running on Cloudflare's edge runtime for high availability and reliability. We used iMessage and Poke as our main agentic interface. We use a custom voice agent deployed on LiveKit Cloud, along with Telnyx to route the agent to dial a real phone number. After the call is done, the transcript is saved to Cloudflare KV, so that Poke can access it and summarize the call through our MCP.

Challenges we ran into

On the first day, we had many troubles getting our telephony system and voice agent to work properly: constant vague error messages were throwing us off. We had to consult with engineers at the sponsor booths to debug what was wrong, but we eventually got through with real phone calls placed. We also had issues with our transcript and storage saving, but we were also able to fix that too.

Accomplishments that we're proud of

We're proud of having a working, full end-to-end voice agent work on our behalf, with 10 hours left to spare! After working out the major issues, we were able to work efficiently and get our project done, with enough time to sleep well!

What we learned

We learned how agentic AI interfaces are created, how the Model Context Protocol works and the power it has for AI tools, how to orchestrate agents in realtime, and how to fine tune and prompt agents for their best use case.