Speelka Preview

I really like how https://speelka.korchasa.dev is turning out so far. For example, because it simultaneously implements an MCP server and knows how to work with MCP servers, we get the ability to build agent hierarchies out of the box. For instance, with just two configurations, you can teach Claude Desktop or any other client to collect a news digest from specified sources (any for which there is an MCP server) in the format you want to receive it. A simple and clear working scheme…

Why is it needed at all:

• building hierarchies of agents with different specializations
• splitting large complex algorithms in prompts into separate small and highly specialized ones
• processing different parts of large agent tasks with different tools, different LLMs with different settings
• being able to expand, narrow, or change the response structure of MCP servers
• cheaply switching between “real” and LLM implementations of a particular tool
• calculating the total budget and telemetry for the entire task implementation chain (planned for the near future)