Skip to main content
The LLM node can do more than read and write text. Bind one or more connected apps — Gmail, Jira, Slack, and 2,700+ more — to the node, and the AI can use those apps while the step runs: look up an email thread, create a Jira ticket, post a Slack update, then write its answer. Think of it as a mini agent inside your workflow. You decide which apps and which actions it may touch; the model decides when to call them and with what inputs.
LLM node config panel with the Integrations section expanded

The Integrations section in the LLM config panel, right after Built-in Skills. The summary row shows what's bound; Configure opens the panel.

​
How it works

Every action you enable becomes a tool the model may decide to call — the same way actors in a Chat & Voice Agent use capabilities. During a single LLM step the model can call several tools, in any order, as many times as the task needs, before producing its final response. Each action call runs on the connection you bound — the model acts as that account.

​
LLM Integrations vs Integration Action

Both can send an email. The difference is who decides.
If you know exactly which action to run and what to pass it, use Integration Action — it’s predictable and easier to debug. Reach for LLM integrations when the step needs judgment: choosing between actions, chaining several, or extracting the inputs from messy text.
Repeating the same calls over many records — one message per row, one contact per signup? Bind the apps to an Execute Code node instead. Your code runs the calls back to back, with no model turn — and no model cost — per call.

​
Set up integrations

1

Open the Integrations section

In the LLM node’s config panel, expand Integrations (right after Built-in Skills). The badge shows how many apps are already connected. Click Configure on the summary row to open the integrations panel.
2

Add an app

Click Add Integration and browse or search the catalog. Pick the app the step needs. Added apps appear as cards showing the app, its provider, how many actions are enabled, and the connection’s health.
Integrations panel with Gmail and Jira bound to the node

The integrations panel. Each card shows the enabled action count, connection health, and whether the model has guidance on when to use the app.

3

Choose a connection

Pick one of your existing connections or click Connect to authorize a new account. Connections are shared workspace-wide — see Integrations for how to manage them.
4

Pick which actions to enable

Tick only the actions this step actually needs. Use All / None to bulk-toggle. You need at least one enabled action to save.
5

Add "When to use?" notes

For each enabled action, click When to use? and write a short note telling the model when it should run that action. The model reads these notes every time it decides what to do next — they are the single biggest lever for reliable behaviour.
Gmail action selection with a When to use note expanded

Configuring Gmail: pick the connection, tick the actions, and add a "When to use?" note to guide the model.

6

Save

Click Save and then save the node. Enabled actions are now available to the model on every run of this step.
Enable only what the step needs. The model decides on its own when to call an enabled action, and every call runs for real on the connected account. If you enable “all of Gmail”, a prompt that mentions deleting can end with deleted emails. Narrow scope is the safety mechanism.

​
Custom API calls

When no action covers what the step needs, set Custom API calls in the app’s panel to Read only or Read and write. The model can then call any part of the app’s API with the same connection. It is Off by default; see Custom actions for what each setting allows.

​
Connection health

Each integration card shows the state of its connection so a broken binding never fails silently: At run time, an integration whose connection is missing or disabled is skipped — the node still runs, and the skipped apps are reported in the node’s skippedIntegrations output.

​
Guiding the model

An enabled action is an ability, not a command. The card badges tell you whether the model has guidance:
  • Usage notes — at least one enabled action has a “When to use?” note. Recommended.
  • On user request only — no notes yet; the model will only reach for the app when the prompt explicitly asks for it. Mention the app in your prompt or add notes.
You can combine both: describe the overall task in the node’s prompt, and put per-action rules (“never email anyone except the customer on the ticket”) in the notes.

​
Cost

The LLM node itself is free to run. The final cost depends on how many times the model invokes each enabled built-in skill or integration action — each call is billed individually. A step that decides not to call any tools costs nothing extra.

LLM

The node this feature lives on — instructions, models, outputs.

Integration Action

Run a single, fixed app action deterministically.

Integrations

Connect accounts and manage connections workspace-wide.

Actor capabilities

The same concept for Chat & Voice Agent actors.

Execute Code integrations

Call the same apps from Python, in loops and batches.