The premise is honest and narrow: a 4B model fails structured reasoning not because it lacks the steps but because nothing makes it take them, and enforcing the structure externally recovers some of that. If you are running a small local model — for cost, for privacy, for offline work — that is a real gain. If you are already calling a frontier model, the scaffolding adds latency to something the model does natively. Worth deciding which situation you are in before attaching it, because the tool cannot tell you and will happily wrap either.
A single-tool reasoning harness: it takes a problem and enforces an agentic chain-of-thought with adaptive parsing, aimed at making small models work methodically rather than answering from pattern.
- solve_problem enforces the step-by-step structure rather than requesting it, which is the difference between a prompt that suggests reasoning and a harness that requires it.
- The README's own example is SEND + MORE = MONEY — a cryptarithmetic puzzle with eight unique digits, five columns and four carries — which a bare 4B model does not solve.
MODEL, naming the model to drive, and a current Node runtime.
One command — npm install @slbdn/cotforce-mcp