Nodes¶

Config tabs¶

Select a node and the Config panel shows its type, id and Label, then the same three tabs for every node, always in this order:

A tab the node does not use stays in its place, greyed out. The panel keeps the tab you chose when you select another node, so you can open Data and click through the graph to see where each node writes. A red dot on a tab means something there needs attention, such as a missing output field.

Execution settings¶

LLM, Agent, Decision, Retriever, Tool, MCP tool and Human nodes have these settings in their Execution tab:

Blank means off.

Start and End¶

Every graph has exactly one Start. End nodes close a path. Neither can be deleted or copied. Start receives the run input as the initial state.

LLM¶

While a run is going, token deltas arrive as node_token events and show in the Logs row and the Output tab. They are not stored in run history; the final text is. Provider failures (401, timeout and so on) become a node error event, never a server crash.

Fallback model¶

In an LLM node's Settings, enable Use a fallback model and choose its credential and model. The pair must differ from the primary. The alternate credential is checked using the workflow owner's access before either model is called; only chat-model credentials are accepted.

The primary is tried first. An unreachable provider, missing model or bad provider response can select the alternate once. Authentication/configuration and credential-access errors stop the attempt. Cancellation or an expired node timeout never starts the alternate. Node retry settings repeat the whole policy, so keep their attempt count in mind when estimating usage.

With fallback enabled, answer text is held until a model succeeds, then emitted as a completed chunk. Ordinary LLM nodes still stream live. The trace's model_attempts records provider/model, status and error code; a failed call can still incur provider charges even if it reports no usage. Mock runs remain offline and do not resolve credentials. Code exports currently reject a workflow with a configured fallback policy (MODEL_POLICY_EXPORT_UNAVAILABLE) rather than silently omitting it. Portable templates clear the fallback credential; select your own credential after import.

Agent¶

A task prompt, a list of tools (MCP, HTTP, built-in function, knowledge base search) and a maximum number of rounds. The model calls tools until it answers; a tool can need a person's approval before each call. See Agent node.

Decision (Jev)¶

Asks TypeSafe's Jev typed questions about text (a template; an exact {{state.x}} sends the value with its type). Each question has an id, instructions and a type:

A question's field receives its answer, so a Condition can branch on a choice. output_field (an object) receives every answer with its probabilities and confidence. The node uses a TypeSafe credential, then TYPESAFE_API_KEY, then offline sample answers (a choice picks the first option named in the text). Usage is priced at $0.042 per million input tokens; output tokens are free. A recovered run reuses its answers.

Retriever¶

Tool¶

A Tool node has three modes.

MCP tool¶

server_id, tool, arguments and output_field. Argument strings are templates; an argument that is exactly one {{state.x}} keeps the value's type, so {{state.items | length}} passes a number. See MCP servers.

Send email¶

Sends one e-mail through an SMTP credential (Credentials → SMTP: host, port, STARTTLS / SSL / none, user name, sender; the password is stored encrypted).

• to, cc, bcc and reply_to are templates with comma-separated addresses; one {{state.x}} may also hold a list. At most 50 recipients.
• subject and body are templates. With format HTML, values from the state are escaped and a plain-text part is added.
• output_field (optional, an object) gets {sent, dry_run, to, cc, bcc, subject, message_id, refused}.
• Without a credential the node is a dry run: nothing is sent, and the result holds the message with a text preview.
• A recovered run never sends again, and the Message-ID comes from the node's idempotency key.
• Private or loopback SMTP hosts need STUDIO_SMTP_ALLOW_PRIVATE=1 (for example Mailpit on localhost:1025).
• Each workflow owner's nodes reach at most STUDIO_EMAIL_NODE_HOURLY_LIMIT recipients per hour (default 100).

Put a Human node before it to approve a drafted e-mail.

Condition¶

Set Field to a state field. Outgoing edge labels are matched against its value:

1. exactly;
2. case-insensitively, after trimming spaces and punctuation;
3. as a whole word in the value, where the longest label wins.

Label order does not matter: Disapprove. picks disapprove, not approve. Two whole-word labels of the same length are ambiguous. The optional default edge is the fallback for no match or an ambiguous one.

A Condition is the only node that may have an edge back to itself, and the only way (with For each) to close a loop.

Parallel branches¶

Several outgoing edges from a node that is not a Condition run in the same superstep. A state field written by more than one branch needs a reducer (append, merge or add); a missing reducer fails validation. Use a Join to wait for all branches.

For each¶

It needs two outgoing edges labelled each and done; the editor labels the first edge you draw each and the second done. The each branch is the loop body and must lead back to the For each node. Every visit writes the next item and follows each; after the last item it follows done.

Items run one at a time, so Human pauses, recovery and idempotency keys work per item. Collect results in a list field with the append reducer. Loops can nest; the inner list starts again for every outer item. Loop steps do not use up STUDIO_RECURSION_LIMIT, which only bounds the other loops.

Human¶

prompt (a template) and output_field. The run waits in status waiting until someone answers Approve (approved), Reject (rejected) or free text. See Human approval. A model can review the request first, as a suggestion or answering on its own: see AI review of approvals.

Join¶

Waits for every incoming branch, then runs once. It needs at least two incoming edges and cannot take an edge straight from a Condition.

Subgraph¶

workflow_id of one of your saved workflows. State is shared by field name. Inner nodes appear in the trace as Subgraph › node (id sub/node). Cycles, a missing or imported child, and nesting deeper than five levels fail on save and on run (WORKFLOW_INVALID).

Code (imported)¶

Read-only nodes produced by repo import. The right panel shows the original source. Editing the canvas does not change the imported file.