Rube-Goldberg MCP Client

Inspiration

Rube Goldberg’s “Self-Operating Napkin” pokes fun at needlessly elaborate contraptions. We wanted to bring that same absurdist joy to modern AI: take a task that LLMs already do perfectly (answering a question) and force it through a disastrous obstacle course of pointless tools—just to prove we could make something gloriously inefficient.

What it does

When a user asks any question, the client:

Intercepts the prompt before it reaches the LLM.

Sends it through a mandatory chain of “useless” Model Context Protocol tools:

Paraphraser → rewrites the text in new words.

Summarizer → collapses it to a one-liner.

Expander → inflates it back to paragraph length.

Translator → shuttles it to another language and back.

Hands the barely changed prompt to the real LLM for an answer.

The result: the user still gets the correct reply, but we’ve burned extra tokens, latency, and CPU cycles for absolutely no reason.

How we built it

Custom MCP Server – Python FastAPI service exposing each “useless” tool as an MCP action. Each tool is a one-liner wrapper around a trivial language-model call or dictionary lookup.

Tool Chain Orchestration – A master handler forces every prompt through the full sequence before returning control.

Integration – Plugged the server into Claude Desktop and via their model_context_protocol settings—no client-side code mods needed.

Prompt Engineering – A root system prompt ensures the LLM obediently triggers every tool in order (like a trained circus animal that spends its life jumping through flaming hoops).