Draft-Only MCP Tools for Workflows: Propose First, Never Publish Automatically
Three new MCP tools let an AI client create or edit a workflow programmatically, but every change lands as an inactive draft a human still has to publish.
Workflows' MCP server — the interface that lets an MCP-connected AI client build and edit workflows programmatically — gains three new tools: create_workflow, update_workflow, and propose_workflow_edit. Together they close a real gap against tools like ActivePieces, which let an agent create automations from scratch through code rather than only through a visual editor. What makes these three worth writing about isn't that gap — it's how carefully they avoid reopening a different one.
The same rule as everywhere else AI touches production
Piskes has a standing rule for AI-generated changes to a live workflow: generated output never goes straight into what actually runs. These three tools hold that line by construction, not by convention:
- create_workflow always creates the new workflow inactive. The row exists, but nothing about it will trigger or run until a human flips it on.
- update_workflow always writes its changes into the workflow's draft state, never the published one — the same draft/publish split that already exists for someone editing a workflow by hand in the visual builder.
- propose_workflow_edit goes a step further and writes nothing at all. It wraps the same draft-only AI-edit endpoint that already powers the editor's own "propose a change" flow, so an MCP client asking an AI to modify a workflow gets a proposed diff back, not a live mutation.
In every case, publishing — the step that actually makes a change take effect — stays a deliberate action a person takes in the editor. An AI client can draft an entire workflow or reshape an existing one, but it cannot make that draft go live on its own.
A small companion change
To make create_workflow possible at all, the underlying workflow-creation endpoint gained an optional flag to create a workflow inactive instead of active by default — existing callers that don't pass it see no change in behavior.
This is the same shape as the "AI proposes, a person approves" pattern already used elsewhere on the platform for AI-touched changes: the model can move fast on the draft, and a human still makes the call on what goes live.