Workspace

A canvas where you draw the steps yourself. The assistant decides how to do a job; a workflow does exactly what you drew, every time.

Open Workspace from the left sidebar (/user/workspace).

Why draw it

Ask the assistant to do something and it works out its own approach — which is what you want when the job needs judgement, and not what you want when it needs to be identical every run.

A workflow is the second case. You place nodes, wire them together, and that is the sequence. Same shape, same order, every time.

[Gmail] → [Filter: urgent?] → [AI: summarise] → [Slack]

Building

Drag nodes from the palette, then drag from one node's output circle to another's input circle to wire them. A node with nothing wired into an input it needs will say so when you run it.

Run all saves the canvas and executes it. Nodes light up as the run moves through them: green for done, red for the one that failed, dimmed for anything a branch chose not to take.

The palette

Sources — Text, Image, Audio, Video, Trigger

Transforms — LLM, Combine text, JSON parse, Text → Image

Control flow — Branch, Loop, Join, Delay, Subflow

Actions — Slack, Email, Webhook, HTTP request, Agent step

Apps — every tool your connected accounts expose, roughly a hundred of them: GitHub, Notion, Airtable, Sheets, Drive, Gmail, Stripe, Shopify, HubSpot, Jira, Trello and the rest. The tab only shows what you have actually connected, so you are never offered a node that can only fail.

The ones worth understanding

Branch sends the run down one of two paths. Both sides exist on the canvas; only one runs. Everything hanging off the untaken side is skipped — dimmed, not red — and the run still counts as successful. That is what makes a branch a branch rather than a failure.

Loop and Join work as a pair. Loop takes a list and runs the nodes between it and the Join once per item; Join collects the results back into one. A sheet that gained thirty rows produces thirty passes and one collected result.

Subflow runs another of your workflows as a single node. Its inputs become that workflow's input, and its result comes back. Useful when several workflows share a middle section.

Agent step is the escape hatch: one step handed to the assistant, which picks its own tools. Some steps genuinely cannot be drawn — "research this company and tell me whether they are worth calling" is a different sequence of lookups every time. Everything either side stays deterministic.

LLM takes a model name if you want a specific one — claude-…, gemini-…, glm-…. Leave it empty for the platform default.

When a run fails

The run stops at the failing node and says which one and why. Every node's inputs, output, duration and error are kept, so a twelve-node graph that broke tells you exactly where.

Running it again resumes from the node that failed rather than repeating the eight before it. For a workflow that already sent email, that is not a cosmetic difference.

Nodes that fail for a transient reason — a model call, an HTTP request — retry on their own. Nodes with side effects, like sending an email, never retry: a duplicate send is worse than a failure.

Credits

A workflow costs credits per node that actually executed. A branch that skipped half your canvas is not charged for the half it skipped.

Credits are checked before the run starts, so you are told you cannot afford it rather than finding out halfway with half the emails already sent. A run that fails partway still pays for the nodes that got through — they did real work.

Results

Terminal outputs — the nodes nothing else consumes — are filed in Artifacts when a run succeeds. Intermediate values stay on the run's node records.

Workspace or Automations?

Flexible versus predictable. Most people use both.