Nodes

A workflow’s nodes are its steps and its edges connect them. Every node has an id, a type, a config, and an optional label that’s shown in the graph. This page lists each node type’s config fields. The authoritative definition is the JSON Schema at https://api.prod.major.build/public/workflow.schema.json.

Use underscores, not hyphens, in node ids. A hyphenated id can’t be referenced from a $expr expression.

agent_call

Runs an agent with a prompt.

Field	Required	Description
`agent_id`	Yes	The agent to run.
`prompt`	Yes	The instruction sent to the agent. Supports `{{ }}` interpolation.
`output_schema`	No	A JSON Schema for the agent’s answer. With it, the node’s output has exactly that shape. Without it, the output is `{ "result": "<final message text>" }`.
`approval_channel`	No	`{ "type": "slack", "channel_id": "...", "channel_name": "..." }`. Sends the agent’s tool approval requests to a Slack channel instead of the app.

app_call

Calls an endpoint on one of your deployed apps.

Field	Required	Description
`app_id`	Yes	The app to call.
`method`	Yes	`GET`, `POST`, `PUT`, `PATCH`, or `DELETE`.
`path`	Yes	The endpoint path, for example `/api/orders`.
`input`	No	An object of values. Sent as query parameters for `GET` and `DELETE`, and as a JSON body otherwise.
`max_retries`	No	How many times to retry a failed response, from `0` to `10`. Defaults to no retries.
`retry_on`	No	Which failed responses to retry: `4xx` (default) retries only 4xx responses, `any` retries any non-2xx response.

The node’s output is the response body. A body that isn’t a JSON object is wrapped as { "body": ... }. Retries back off exponentially, starting at 2 seconds and capped at 30 seconds between attempts. A call where the app never answers isn’t retried. In the editor, set Max retries and Retry on in the node inspector. In the workflow editor, pick the Endpoint from the endpoints Major extracted from the app’s deployed API, or click Enter manually to type the method and path. If the endpoint’s path has a parameter such as {id}, fill it in under Path. Edit input in the node’s Request body field, or its Query params field for GET and DELETE.

router

Branches the run on state.

Field	Required	Description
`branches`	Yes	A list of `{ "when": { "$expr": "..." }, "to": ["node_id", ...], "label": "..." }`.
`default`	Yes	Node ids to run when no branch matches.
`default_label`	No	Label for the default path in the graph.

Branches are checked top to bottom and the first one whose when is true wins. A condition that evaluates to null doesn’t match. Every target in the winning to list runs in parallel. The router must have exactly one outgoing edge per target named in its branches and default.

for_each

Runs a loop body once per item in a list.

Field	Required	Description
`input.iterable`	Yes	The list to loop over, usually a `$state` or `$expr` reference.
`mode`	No	`sequential` (default) or `parallel`.
`max_concurrency`	No	Upper bound on parallel iterations.
`max_iterations`	No	Upper bound on the number of iterations.
`collect_results`	No	Defaults to `true`. Set to `false` to skip collecting results.

A for_each has two outgoing edges: one with "loop": "body" into the loop body, and one with "loop": "exit" that runs after the last iteration. Every path from the body must lead back to the for_each. Loop bodies are the only cycles a workflow may have. Inside the body, <for_each_id>.item and <for_each_id>.index are the current item and its position. After the exit edge, <for_each_id>.results holds each iteration’s final output, in item order.

join

Waits for parallel branches to meet.

Field	Required	Description
`mode`	Yes	`all`. The join releases once every inbound edge has arrived.

Every edge into a join must come from the same fan-out (a router branch with several targets, for example), with no other join in between. Keep each parallel leg a simple line into the join.

wait

Pauses the run.

Field	Required	Description
`duration_seconds`	Yes	How long to wait, in seconds.

jev_noul, jev_choice, and jev_score

Ask Jev (TypeSafe’s judgment model) a question about some text, then branch on the answer with a router. Use them to filter, classify, or rate content. In the editor they’re called Jev yes/no, Jev choice, and Jev score. All three take a text field: the text to judge. It supports {{ }} interpolation, so it’s usually built from run state, for example "{{trigger.input.email_body}}".

jev_noul

Asks one or more yes/no questions.

Field	Required	Description
`text`	Yes	The text to judge.
`questions`	Yes	An object of `{ "<question_id>": "<yes/no question>" }`, with at least one entry. Question ids must be letters, digits, and underscores, and can’t start with a digit.

The output is each question’s yes-probability, from 0 to 1, under its id. For a node judge with a question is_urgent, branch with judge.is_urgent > 0.5.

jev_choice

Picks one label from a list of options.

Field	Required	Description
`text`	Yes	The text to judge.
`question`	Yes	The question to answer, for example `"What is this email about?"`.
`options`	Yes	An object of `{ "<label>": "<description>" }`, with at least two labels. Set a description to `null` to leave it out.

The output is { "choice": "<label>", "confidence": ..., "probabilities": { "<label>": ... } }. Branch with, for example, classify.choice == "billing".

{
  "id": "classify",
  "type": "jev_choice",
  "config": {
    "text": "{{trigger.input.email_body}}",
    "question": "What is this email about?",
    "options": {
      "billing": "Payments, invoices, refunds",
      "bug": "Something is broken",
      "other": null
    }
  }
}

jev_score

Rates the text on an ordered scale.

Field	Required	Description
`text`	Yes	The text to judge.
`question`	Yes	The question to answer, for example `"How urgent is this?"`.
`levels`	Yes	A list of at least two level descriptions, lowest first. The level at position `i` is score `i`. Use `null` for a level with no description.

The output is { "score": ..., "confidence": ..., "probabilities": ... }. score is the expected score from 0 to the last level’s position, so it can fall between levels. Branch with, for example, rate.score >= 1.5.

human_approval

Asks a person to choose in Slack. Requires the Major Slack integration.

Field	Required	Description
`channel`	Yes	`{ "type": "slack", "channel_id": "..." }`.
`message`	Yes	The message to post. Supports `{{ }}` interpolation.
`options`	Yes	Buttons, each `{ "value": "...", "label": "...", "style": "..." }`. `value` is required; `style` is `primary`, `danger`, or `neutral`.
`timeout_seconds`	No	How long to wait for a click.
`on_timeout`	With a timeout	The option `value` to use if no one answers in time. Required when `timeout_seconds` is set.
`reminder_interval_seconds`	No	Post a reminder at this interval while waiting.

The output is { "choice": "<value>", "responded_by": "<user id>" }, or { "choice": "<on_timeout value>", "timed_out": true } on timeout. A human_approval has one outgoing edge, which should lead to a router that branches on <node_id>.choice. Only Major members with access to the workflow can answer.

end

Finishes the run.

Field	Required	Description
`status`	No	`success` (default) or `failure`.
`output`	No	An object of values to record as the run’s output.

An end node has no outgoing edges.

Edges

Each edge is { "id": "...", "from": "node_id", "to": "node_id" }, with an optional label shown in the graph. Only for_each out-edges carry "loop": "body" or "loop": "exit".

Graph rules

Saving checks the definition against these rules, and errors name the path that failed:

At most 50 nodes, and the file is at most 256 KB.
Node ids and edge ids are unique. trigger is reserved and can’t be a node id.
entry and every edge’s from and to name an existing node.
Only router and for_each fan out. Every other node has exactly one outgoing edge, and end nodes have none.
A node can have several inbound edges only if at most one of them can be active at a time. Parallel paths must meet at a join first.
Every node is reachable from entry, and every path reaches an end node or loops back to a for_each.

Get Started

Learn

Build

Settings

Reference

agent_call

app_call

router

for_each

join

wait

jev_noul, jev_choice, and jev_score

jev_noul

jev_choice

jev_score

human_approval

end

Edges

Graph rules

Next steps

State and input

Triggers