🧩Workflows

Node reference

The node library ships with around two dozen node types across five groups. Mix them freely to build anything from a two-step notification to a branching, AI-driven pipeline.

ℹ️
Two notes before you start

The Web Search node (and an agent's web tool) query the Brave Search API, governed by a per-organization search budget. The Code and HTTP Request nodes can be disabled per workspace by an admin, so they may not appear in every workspace's library.

Triggers

Start a workflow. The first node must always be a trigger.

NodeWhat it does
▶️ ManualRun on demand from the hub.
🔗 WebhookStart from an inbound HTTP POST.
ScheduleRun on a cron schedule or interval.
🧩 Integration eventReact to an event from a connected tool.
📂 Folder watcherFire when a doc lands in a knowledge folder.
📋 FormStart from a submitted in-app form.

AI

Embed intelligence anywhere in the flow.

NodeWhat it does
LLMA single prompted model call. Free text or structured JSON.
🤖 AgentRun a full agent with tools, knowledge, and sub-agents.
🌐 Web SearchSearch the web via the Brave Search API.
🛡️ GuardrailsCheck or constrain content against rules before it proceeds.
🎨 Image GenerationGenerate an image from a prompt.
🔎 File SearchRetrieve relevant passages from knowledge / files.

Control flow

Branch, repeat, and fan out.

NodeWhat it does
🔀 ConditionBranch on a condition — separate true / false paths.
🔁 LoopIterate over a list, running the body for each item.
ParallelRun multiple branches at the same time and join.

Actions

Do work and move data.

NodeWhat it does
💻 CodeRun a custom code snippet to transform data.
🌍 HTTP RequestCall any HTTP API; parses JSON responses.
🧩 Integration ActionPerform an action in a connected tool (create issue, send email…).
📚 Knowledge FolderRead from or write a document to a knowledge folder.
🗃️ TableDescribe, validate, and materialise tabular data.
🔢 Tabular QueryQuery tabular data with SQL.
📎 Insert ContextFold an upstream node's output into the context.
📌 Publish ResultPublish a deterministic result for downstream use.
DelayPause the run for a set duration.
🔔 NotificationSend a notification to your team.

Terminal

End the flow or pause for a person.

NodeWhat it does
📤 OutputReturn the final result and end the run.
Human ApprovalPause and wait for a person to approve or reject.

Configuring AI nodes

AI nodes see the output of upstream nodes and can return free text or structured JSON for downstream steps. The LLM node is a single prompted model call; the Agent node runs a full agent with tools, knowledge retrieval, and optional sub-agents. Both expose the same core settings:

SettingDescription
ModelPick any model an admin has enabled. Each node chooses independently across the six providers.
System promptInstructions for the AI. Insert upstream values with {{ }} context references.
Knowledge foldersAttach folders the node searches during the run.
ToolsEnable web search, integration actions, and more for the node to call.
Structured outputDefine a JSON schema so downstream nodes can reliably parse the result.
Temperature / top-p / max tokensTune creativity, sampling, and response length.

For a node that delegates to one of your saved Agents, select the agent and it supplies its own model, prompt, tools, and knowledge — you don't re-configure them on the node.

Control flow in practice

Control-flow nodes shape how execution moves through the graph. A Condition node splits into "true" and "false" paths so you can route work based on data. A Loop node iterates over a list, running its body once per item. A Parallel node fans out into branches that run at the same time and then joins their results — useful when several independent steps can happen concurrently.

Ending a flow — and pausing for a human

An Output node returns the final result and ends the run. A Human Approval node instead pauses the run and waits for a person to approve or reject before continuing — covered in detail under Running & approvals.

ℹ️
Budget-aware by design

Every AI node and integration action draws on your organization's budget, and when a budget is reached the relevant capability pauses automatically so costs stay predictable. Every numeric cap has a hard minimum, and 0 means disabled, never unlimited.