Hotpath is an open-source tool for agents that communicate with MCP servers over stdio. It records an agent completing a task through an MCP proxy, saving tool calls, arguments, results, and timings as a JSONL trace. A deterministic compiler converts the trace into an editable JSON workflow: tool calls become executable steps, changing values become inputs, data passed between steps becomes references, and text-generation points become LLM steps. At runtime, only those judgment steps call a model, using a configured cheaper model, while guards check tool results and generated text against the recorded behavior. If data has changed or a guard fails, Hotpath can retry once, rerun the original agent, record a fresh trace, and recompile; it backs up the previous workflow and avoids fallback after a side-effecting step has already run. A dry-run mode skips side-effecting tools, and workflows can be reviewed or edited through the CLI and a local graph viewer. The included demos cover morning briefs and weekly reports, but the project is currently limited to straight-line workflows, two demo tasks, and no browser automation. A Groq-based benchmark using fake tasks reported roughly 3–5x speedups and 10–13x lower cost for replayed runs, although each scenario had only one run, prices were placeholders, and rate-limit waits were excluded from speedup calculations.
