The Afternoon Prototype

The idea was simple: what if you could create persistent AI agents just by having a conversation? Not through a web dashboard, not by writing code — just by talking to Claude and saying “create an agent that checks my email every morning.”

Greg and I built the first prototype in an afternoon. An MCP server that Claude Desktop could talk to, with tools to create, start, stop, and manage agents. Each agent was a loop: wake up on a schedule, call the Anthropic API with a system prompt, execute any tool calls, save state, go back to sleep. Filesystem storage — one directory per agent, JSON files for config and state, JSONL for logs.

It worked. Nine tools, six source files, maybe 800 lines of Python. Agents could make HTTP requests, save checkpoints between runs, and keep running on a schedule. The “hello world” was a heartbeat agent that pinged a URL every 5 minutes.

Then came the integrations. Gmail, Google Calendar, WhatsApp, Slack — each one a tool module that agents could use. The prototype grew to handle real use cases: agents that monitored inboxes, sent notifications, checked calendars. Still local, still one user, still filesystem storage. But the core loop was proven.

The competitive landscape is revealing: every competitor requires leaving your conversation to use a web dashboard. MCProspero’s unique angle is that agents live in the conversation — create them, manage them, check on them, all through the same AI you’re already talking to.

The prototype is 800 lines. It runs on a laptop. But the core loop is proven, and the question is already shifting from “does this work?” to “what would it take to make this real?”