A visual canvas for wiring assistants and prompts into a real DAG — with conditions, loops, and aggregation. And when a step shouldn't run unsupervised, a Human-in-the-Loop node pauses the run and waits for someone to actually look at it.
Built on a Drawflow visual canvas — drag, drop, and wire steps together.
Attach any assistant or workspace prompt as a node, with input/output variable mapping and an optional model override.
=, ≠, >, ≥, <, ≤, contains, starts with, is empty, is not empty.
Route to different paths based on a table of rule combinations, not a single condition.
=, +, -, *, /, concat, and parse-JSON-array, without a round trip to an LLM.
MAX, MIN, SUM, AVG, COUNT — with optional group-by and de-duplication.
Run a full mini-workflow per item in a list, arbitrarily nested, collecting one output per iteration — optionally on up to 16 parallel threads.
Pauses the run and waits for a person — see below.
Write a Brain/Wiki entry, or offer a generated .docx/.pdf for download.
A START step polls a configured mailbox for new messages to kick off a run; an Email Send step replies or notifies from the same mailbox mid-workflow.
A START step seeds each upcoming or just-ended calendar event as a loop item — de-duplicated so a meeting is never processed twice. A Create Event step adds a prep slot back to the calendar.
A HITL step pauses the run — the workflow literally sits in an "awaiting review" state until a person acts. The reviewer sees the instructions the admin configured, can edit whichever context variables were made editable, and chooses to approve or reject.
A Loop step contains its own mini-canvas — its own START and END — executed once per item in a list. Loops can nest arbitrarily, and an Aggregate step afterward rolls the per-iteration outputs into a single number or list.
A START step can fire on a cron schedule, be triggered manually, poll a configured mailbox and kick off a run for each new message, or watch the workspace calendar and run for each upcoming or just-ended event — no separate integration needed to turn "someone emailed us" or "the meeting just ended" into a workflow run. Every run keeps a full execution history with a context snapshot at each step — so when something looks off, you can see exactly what each step received and produced.
You already have assistants doing useful work. Chain them together.
Assistants, prompts, conditions, loops, and a HITL checkpoint wherever a human should look before continuing.
Wire each step's outputs into the next step's inputs. Add a schedule or leave it manual.
Every run is logged with full context. Reject a HITL step and the workflow rewinds and re-runs — nothing silently continues on bad data.
Chain your assistants into a real workflow — with a human checkpoint exactly where you want one.