The builder

The builder is where a workflow is created. On the left is the column Build, a chat that creates and changes the steps for you. In the middle is the canvas with one card per step.

Open the builder

There are three ways into the builder:

• From the chat: in an empty conversation, click Create workflow below the input field and describe the process. The chat shows the card Build workflow. Build opens the builder, and the chat there starts building at once.
• Empty: on the page Agents, choose Create empty workflow in the menu with the three dots. You get a trigger Form and a step New step with an empty instruction.
• From a template: on the page Agents, open Templates and ideas. The chat then asks for what the template cannot bring along, such as recipient, folder or mailbox.

You find ideas on the templates page and in the chat. Clicking an idea opens a dialog. Adjust the description in the field Order to the builder, then click Build.

Build in the chat

Write in the column Build what should happen. The chat changes the cards on the canvas directly.

On an empty canvas, the chat first works out which tools it needs and which fixed values only you can name, such as recipient, folder or channel. Then the card Build plan shows the planned steps as a checklist and ticks off each step as it is created. Afterwards lowcloud checks the process, and the chat fixes errors in a second pass. If something remains open, the round ends with "Workflow is not yet executable: …".

At the end, the chat writes whether the workflow is executable. If fixed values are missing, it names every one of them. You answer in the chat or enter the value directly on the step. For folders, workbooks and tables in SharePoint or OneDrive, and for a team, channel or chat in Microsoft Teams, the chat shows a picker with the button Apply instead of a question.

• Above each answer is a collapsed line with what the chat did. A round without a change is called Thought.
• The microphone dictates into the input field.
• If a round fails, the card That did not work appears above the input field, with the reason and the button Again. The input field remains usable.
• The model of the chat is always set to Automatic and cannot be changed. You choose the model of an individual step freely.
• The chat cannot connect an account, bind a mailbox or create a skill. It says what is missing and points to the trigger or to the connections.
• The chat history is saved with the workflow and loaded again.

Change a workflow in the chat

Write the change into the chat, for example "make the answer shorter" or "send this via Teams instead of email". You can also ask a question, for example "What does this workflow do?". A question changes nothing in the workflow, and the answer says so.

If the connection for a service is missing, the chat shows the card Connection missing with a button Connect per service. Without a connection, the step fails in the run.

If a card is open or a run is selected in Runs, a reference appears above the input field, for example Reference to the step …, Reference to the trigger … or Reference to run #7. As long as it is there, "make this shorter" means this step. The × next to it switches the reference off for the next messages. If you open another card, it comes back.

Edit steps by hand

Click a card to open its dialog. The chat next to it stays usable and knows which step is open. With the arrows at the left and right of the dialog, you go to the previous or next step. Escape, the × or a click on the dimmed area closes the dialog. Which fields a step has is described under Steps.

The row Note above the tiles is for you and the chat. The agent never reads it in the run. The chat reads it and writes notes itself.

Undo (⌘Z) and Redo (⇧⌘Z) are in the toolbar of the canvas. Next to them, Rearrange steps arranges the cards in a new layout. If you switch workflow, the history for Undo is discarded.

Connect, insert, duplicate and delete cards

• Connect: drag from the handle of a card or of a path row onto another card. You can drag the end of an existing connection to another target.
• Insert: the + on a connection places a step in between. The + on a last card appends a step or connects the card to an existing one. Every path row that ends also has a +. Every + opens the menu New step with Agent, Action and Condition.
• Duplicate: the copy takes over the settings without the paths. Its name gets (Copy).
• Delete: if the step has exactly one target, the gap is bridged. Otherwise the paths end there, with the notice "A path now ends here".

A newly inserted step stays closed and shows a notice until you have written its task.

You cannot delete or duplicate the trigger. You cannot delete the last remaining step. Two paths of one step cannot point at the same target.

Test a workflow

You start a test run in the panel Test run below the trigger card. What the panel needs depends on the trigger: the fields of the start form, a real event from the connected mailbox, or nothing.

1. For an integration trigger, click Load latest events. The panel loads real events from the connected account.
2. Select an event.
3. Click the start button. It is called Start test run, for an event trigger Test with this event, and for a schedule Simulate tick.
4. Answer on the card that stops the run: for Waiting for your approval with Approve or Reject, for The step is asking you in the answer field.
5. Check the result strip of each card.

There are no invented events. Trigger only in the menu next to the start button stops directly after the trigger. During the run, the button shows Stop. For a trigger where nobody is watching, the test runs without questions and notifications.

If the test cannot start, the panel names the reason, for example a red notice on a card.

The result strip shows status, duration, tokens and the number of tool calls. Expanded, it shows output and context. If you change a step, its result stays and is marked From before your change.

Repeat a single step

If a card has a saved intermediate state, it carries a ▶ (This step only). It replays this step against the current canvas without restarting the run. Continue from here then carries on. Everything that came after this step in the run is discarded in the process.

The trigger has no ▶. There, you start a run via the panel Test run.

View runs

Expand Runs at the bottom edge of the canvas. On the left are the runs with status, run, time, duration and tokens. A run against the draft is shown as Draft instead of a version number. On the right is the timeline of the selected run.

A click on a row opens the run. The cards then show what it did, and the next ▶ builds on its state. Replay walks the timeline step by step on the diagram, with Pause, Continue and To the end. What gets marked is the current draft. Steps that no longer exist stay unmarked.

In the run view, you resume or discard a waiting run. You grant or reject open approvals and answer open questions. Below is a second chat, Questions about the run, for questions such as "Why did the agent act this way?". It only reads along and decides nothing.

There are no statistics or success rate in the builder. The figures are on the overview page of the workflow. What a run contains is described under Runs.

Publish a version

What you see in the builder is always the draft. Publish in the header freezes it as the next, immutable version. Nothing starts and nothing stops, and running and waiting runs keep their state. Whether the workflow runs for someone is decided by each person's personal setup, see Publishing and setup.

The display next to the button shows Draft or the version, for example v3. If a version already exists, the button is called Publish changes. If publishing changes the way the workflow starts, a triangle on the button warns and names the change.

If the workflow can be set up personally, the dialog Version 4 published opens afterwards. It asks whether the new version should also run for you. You answer with Set up for me or Update to v4, or you choose Later. Later applies as long as the page stays open.

Reasons for a refusal

lowcloud refuses to publish if

• the process is not executable. The message reads "Workflow cannot be published: …", followed by the findings.
• a step has open fixed values ("… has open fixed values: …").
• a step uses a deleted agent, a deleted custom integration or a deleted MCP server.
• a workflow that the team sees contains a private agent, skill or bucket.
• a borrowed agent is locked for the client.
• the trigger is a webhook ("Webhook triggers are currently not available. The workflow cannot be published.").

You cannot select a webhook in the field Kind of the trigger. The chat can still set one on a newly built workflow. This only becomes apparent on publishing.

Restore an older version

Via the status display in the header you open Versions. The list shows all frozen versions with their date and marks the published one.

• Publish this version makes an older version the published one again. No new version is created.
• Load into draft writes the version into the draft on the canvas. The published version stays unchanged until you publish again.

Build without an account

If you come from the website without signing in, a notice with the button Sign in and keep is shown above the builder. Your workflow is still saved, in a throwaway workspace with one workflow per visitor. Publishing, versions, connecting services and test runs do not work without an account.

Notices on cards

The builder re-checks the process on every keystroke. Red notices above a card block testing and publishing. Yellow notices are warnings and block nothing.

The sparkle next to a notice sends it to the chat as a message. The chat explains it and fixes it if possible.

Save status

You do not have to press a save button. Every change is sent to the server after about one second. You see the state when you click the status display in the header. A click on the row with the state saves immediately.

Limits

Permissions

Anyone who may build (the role admin, or the former role developer) can open the builder. People with the role member land on the workflow list. There they can run workflows, but not build them.

A client workspace is a workspace that a partner created for a client. There, only someone who belongs to the partner's team and may build in the client workspace builds. The client's people do not build there, not even with the role admin. They land on the workflow list. In the partner's own organisation, where the partner manages its clients, nobody builds. See Partner.