Built in less than 48 hours
I discovered OpenAI’s WebMCP Challenge less than 48 hours before the deadline and immediately thought: I have to build for this.
ONE Voice Lab already existed as a broader provider-neutral concept, but its public WebMCP challenge release came together during an intense sprint of roughly a day and a half. That included defining the four-tool browser workflow, creating a visible human-authorization checkpoint, validating the deterministic replay and evidence experience, sanitizing the repository for open-source release, deploying the public application, and producing the demo.
To be fully transparent, not every provider-backed feature is live in this public version. I was not yet comfortable exposing personal API credentials or unfinished production integrations through a public application. Instead, this release focuses on a real and testable interaction model:
- An agent can discover the available tools.
- It can prepare an exact replay plan.
- It cannot authorize its own plan.
- A person must approve the visible action.
- Authorization is consumed after one run.
- The resulting timeline and evidence remain clearly labeled as simulated.
- The experience makes zero provider requests and incurs zero provider spend.
I built this independently, primarily with Codex as my engineering collaborator. I also used OpenAI Prism to turn some of the more abstract systems concepts into structured technical papers before translating them into concrete architecture and product decisions.
That became one of my favorite workflows from this project: if you have a complicated idea living mostly in your head, develop it into a paper in Prism, question and refine the model, and then use Codex to help turn the resulting structure into something people can actually explore.
I entered because I want feedback, not because I believe the work is finished. Feedback from judges, developers, voice AI providers, and people exploring the website will help shape a sustainable roadmap toward a production-ready voice lab built for both humans and agents.
Over the next few months, I am excited to evolve ONE from a deterministic public simulation into an open environment where people can safely connect their own providers, compare real systems, understand cost and performance tradeoffs, and make better voice AI decisions without surrendering human control.
The challenge sprint is finished. The lab is just beginning.
:flying_saucer:
Log in or sign up for Devpost to join the conversation.