/workflow).
The authoring loop
Add steps
Drag step types onto the canvas and connect them with dependencies. Each step needs a unique
id, a type, and type-specific configuration.Wire dependencies
Steps run when their
depends_on predecessors complete. Independent branches run in parallel automatically.Validate
Click Validate to run a server-side dry-run. Validation checks graph structure, dependency cycles, Lua syntax, MCP bindings, and connection overrides — without executing tools.
Import (optional)
Use Import to load a shared graph from JSON — canvas
{ "nodes", "edges" } or backend { "steps": [] }. The review step shows a read-only preview and, for MCP steps, lets you map each server_url to an instance in your workspace before applying to the canvas.Publish
Publish an immutable version when the graph is ready. Runs always execute a published snapshot, not draft edits.
Template variables
Steps pass data downstream through template strings:tool_args, LLM prompt fields, for_each_items, and Lua script inputs. Upstream step results are available under steps.<step-id>.result.
Studio suggests variables as you type — see Template intellisense for search tips, the variable picker, and the Variables reference tab.
Run-start input uses {{input.*}} (from trigger_payload and optional trigger_steps). See Run setup and start input.
Connection overrides
Workflows can override which connection an MCP instance uses at runtime so the same graph works across environments (sandbox vs production credentials). The Console does not expose a dedicated overrides panel today. Set overrides via the API or Platform MCP:Versioning
Each publish creates a new immutable version. You can list versions, roll back runs to a specific version, and compare graphs. Draft edits do not affect in-flight runs.Live streaming
While a run is active, Workflow Studio subscribes to a WebSocket event stream (/v1/workflows/{id}/stream). You see step start, progress, completion, failure, and human-task pause events as they happen.
Autopilot assistance
The Console Chat panel (Autopilot) can scaffold starter workflows, create graphs from natural language, and navigate you to Workflow Studio. Autopilot actions includecreate_workflow, scaffold_starter, and run_workflow.
Next steps
- Template intellisense — inline variable suggestions
- Step types reference — MCP, LLM, human, Lua, and for-each
- Run setup and start input — trigger_payload, saved defaults, API merge flag
- Runs and Command Center — run controls and observability
- Human tasks — approval gates
- JSON schema cookbook — copy-paste graph examples