The trade is explicit: it calls out to a second model, which means another key and another round trip, and in return you cap how many words each reasoning step may use. The two stats tools report how Chain of Draft has actually compared with Chain of Thought on the problems this server has run — check them before deciding it earns a slot.
A reasoning wrapper: your prompt is transformed into a structured drafting format, passed to whichever LLM you have configured, and the response is transformed back into a concise answer.
- chain_of_draft_solve for any reasoning problem, with a domain hint, a choice between the Chain of Draft and Chain of Thought approach, a cap on words per reasoning step, and an adaptive word limit
- Three specialised entry points carrying the same controls: math_solve, code_solve and logic_solve
- analyze_problem_complexity, to size a problem up before deciding how much reasoning to spend on it
- get_performance_stats and get_token_reduction, which report how the two approaches have compared on the problems this server has run
- Bring your own model: Anthropic Claude, OpenAI, Mistral, any Ollama model, or a local chat-completion endpoint
A model of your own and the credentials for it: ANTHROPIC_API_KEY, OPENAI_API_KEY or MISTRAL_API_KEY. For a local model, set MCP_LLM_PROVIDER=ollama with MCP_OLLAMA_MODEL, or MCP_LLM_PROVIDER=custom with MCP_CUSTOM_LLM_ENDPOINT pointing at your own chat-completion API. The repository ships two implementations — Python 3.10+ and Node.js 22+ — and the npm package is mcp-chain-of-draft-prompt-tool.
One command plus a key — npx -y mcp-chain-of-draft-prompt-tool, then supply credentials
