Skip to content

Quickstart

Build a workflow that drafts a reply to a customer message, waits for a person to approve it, and can be called from another service over HTTP.
You will build Start → LLM (draft a reply) → Human (approve it) → End
You will use The editor, the state schema, a run trace, an API token and a schedule
Time About 10 minutes
Video The video on the home page follows the same steps

Before you begin

  • A running studio: ./dev.sh from a clone of the repository (see Install). The UI is at http://localhost:5173.
  • An account: create one with Sign up on the sign-in page. The default editor role is enough for this guide.
  • No model key is needed. Without a credential, LLM nodes answer from a deterministic offline mock that echoes the end of the prompt.

1. Create a blank workflow

  1. On Workflows, select New workflow.
  2. In the gallery, select Blank workflow. The editor opens with a Start node, an End node and one edge between them.
  3. Rename it in the name field of the header, for example Refund reply.

The New workflow gallery

2. Add and connect the nodes

  1. Drag LLM from the Node palette onto the canvas. Scroll the palette down to Human and drag it next to the LLM node.
  2. Select the edge from Start to End and press Del.
  3. Connect the nodes: drag from the right handle of Start to the left handle of LLM, then LLM to Human, then Human to End.
  4. Press Shift+L to tidy the layout (also ⋯ → Tidy up the layout).

Check

The canvas shows four nodes in one row: Start → LLM → Human → End.

The wired graph

3. Configure the LLM node

The state is the typed record that every node reads from and writes to. A new workflow has one field, message, which holds the run input.

  1. Select the LLM node. Its settings open in the Config panel.
  2. Set Label to Draft reply.

  3. Replace the Prompt template with:

    Write a short, polite reply to this customer message: {{state.message}}

    {{state.message}} shows as a tag. Typing {{ lists the state fields; a field that is not in the state is marked as unknown.

  4. Under State schema in the left panel, select Add field and name the new field reply.

  5. Open the Data tab and set Output field to reply. The node writes its answer there.

LLM node settings

4. Configure the Human node

  1. Select the Human node and set Label to Approve reply.

  2. Set Question for the human to:

    Send this reply to the customer?

{{state.reply}}

  3. Add a state field decision and, in the Data tab, set it as the Output field. It receives approved, rejected or a typed answer.

  4. Optional: back in the Settings tab, under Approval channel, set Send the approval request to a connected Slack, Telegram, WhatsApp, e-mail or Teams channel, then its target (for Slack, the channel id). Approvers (e-mails) lists who else may answer. Channels are connected under Settings → Channels; see Approval channels. The quickstart works with Studio only.

    The Human node sending its request to Slack

  5. Select Save (Ctrl+S or Cmd+S). The backend validates the graph and stores it as a new version.

Validation

Save and Run check the graph first: unreachable nodes, references to fields that are not in the state, and edges to missing nodes are listed before anything runs. See The editor.

5. Run it and approve

  1. In the Run tab at the bottom, type into Input (left) and enter:

    {"message": "My order A100 arrived broken. Can I get a refund?"}

  2. Select Run (Ctrl+Enter). The LLM node finishes, and the run stops at the Human node with status waiting.

    The run waiting for approval

  3. Read the question and the drafted reply, then select Approve. The run resumes from its checkpoint and finishes.

Answer from chat or e-mail

With a channel set on the Human node, the same request also goes there: Approve and Reject buttons in Slack and Telegram, two reply buttons in WhatsApp, and links in e-mail and Teams that open the studio. Only the run's owner and the listed approvers can answer, once. Answering in the studio closes the request everywhere. See Human approval.

Check

The logs (right of the input) show Completed with four steps: Start, Draft reply, Approve reply and End.

6. Read the trace

Select Draft reply on the canvas or in the logs. The step opens with Input, Output, State diff and Raw events. The state diff shows reply added by this step.

State diff of the LLM step

7. Call it over HTTP

  1. Open ⋯ → API, enter a token name such as billing-service, and select Create token. The token (wft_…) and the signing secret (whsec_…) are shown once; copy them now.

    A new API token

  2. Start a run from a terminal. Replace <wid> with the workflow id from the address bar and wft_… with your token:

    curl -X POST http://localhost:5173/api/hooks/workflows/<wid>/runs \
  -H "Authorization: Bearer wft_…" -H "Content-Type: application/json" \
  -d '{"input": {"message": "My order A100 arrived broken. Can I get a refund?"}}'

    The response is 202 with a run_id and "status": "queued".

  3. Open Run → View all (reload the page first). The API run has an API badge and status waiting: it stopped at the Human node, like the studio run.

History with the studio run and the API run

8. Put it on a schedule

  1. Open ⋯ → Schedules.
  2. Enter a name and the cron expression 0 9 * * 1-5 (09:00 on weekdays). The dialog previews the next run times in the selected time zone.
  3. Select Add schedule.

What you learned

  • A workflow is a graph of typed nodes over a shared state; nodes read it with {{state.field}} and write one Output field (State and references).
  • Every save is a version, and every run records the version it used (Versions).
  • A Human node pauses a run in waiting and resumes it from a checkpoint; the request can also go to chat or e-mail (Human approval).
  • Each step of a run keeps its input, output and state diff (Runs and traces).
  • A workflow token lets another service start runs, and schedules start them on a timetable (API triggers, Schedules).

Next steps