GPT-4o Realtime API plus computer control plus image generation, all wired into a single voice agent built inside Cursor. Riley Brown's tutorial produces a local AI assistant that listens, searches the web via Exa, controls your mouse and keyboard, and generates thumbnails, assembled through roughly eight incremental prompts with no traditional coding workflow.
The value is in the process, not just the result. Brown builds the agent in named layers, introducing the OpenAI API key, then Exa web search, then computer use mode, then image generation as discrete steps. Each layer is a separate prompt to Cursor, which means the architecture is visible and replicable. The GitHub repo at rbrown101010/rileyjarvis makes the full codebase available.
The final demo at 18:46 is the payoff, but the argument worth reading for is how far commodity APIs have come when composed this way. If you have ever priced out building a voice-controlled desktop agent, the Exa plus GPT-4o Realtime stack Brown assembles here resets your assumptions about complexity and cost.
[WATCH ON YOUTUBE →]