routinekit

Your agent nailed a five-step tool-call sequence, then you need the same thing with different inputs and are about to re-prompt the model to reconstruct every step. RoutineKit removes that repetition. It captures the tool calls you explicitly select, turns them into a typed, parameterized JSON routine with input bindings and success checks, and then replays deterministically with new values. Zero model calls during replay.\n\nThe replay engine rides the same tool-execution path as your host agent, inheriting its scope, approval policies, and cancellation guards rather than spinning up anything new. Drive it from the DeepSeek Harness sidebar panel, the standalone workbench, or a plain stdio MCP server. Local file workflows and WebMCP page calls are both supported, but note: it is a deterministic runner for reviewed linear workflows, not a general computer-use agent or OS sandbox.\n\nSave a routine as a named tool and the owning agent can invoke it directly with new typed inputs. Export the companion SKILL.md and routine.json as a ZIP, and a colleague reviews the JSON, reconnects compatible tools, and approves a fresh run in their own environment. Built for developers working in DeepSeek Harness or MCP clients who want to codify a one-off success into a reusable asset and care that replaying it costs no inference token.