Jan Čurn argues that some problems attributed to MCP come from how agent harnesses load and use toolsAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction.. He describes the context costs of loading many tool definitionsContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. and the risks of putting sensitive or large results into model context.
He compares subagentsA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks., progressive tool discovery and code-based tool use, then argues that a local command-line interface can complement MCP’s standardized remote access.
Čurn demonstrates Apify’s MCPC client, including sessions, tool search and asynchronous tasks. He presents early connector evaluation results as preliminary and invites further contributions.
Watch on YouTube




