AI Metadata

Every action and trigger you build is also an AI tool: agents connected through the MCP server discover piece actions and execute them directly. Two optional fields control how your piece appears to agents: aiMetadata describes the operation in agent terms, and audience controls which surfaces an action shows up in. Both are additive: omitting them leaves the piece behaving exactly as before.

aiMetadata

Available on both actions and triggers:

aiMetadata: {
  description: string,   // optional, agent-oriented description
  idempotent: boolean,   // optional, is repeating the call with the same input safe?
}

description is written for an agent, not for the UI. The regular description stays a short label under the action name in the builder; aiMetadata.description can be a full paragraph that states what the operation does, its notable options and constraints, and how it differs from sibling actions ("Use Send Message To A User for a private DM"). This text feeds the tool search index, so a precise description directly improves whether agents find your action.

idempotent declares whether calling the operation twice with the same input is safe. Reads, upserts, and set-value operations are idempotent; anything that creates, sends, or appends on every call is not. The value is exposed to agents and MCP clients as metadata that informs whether a retry is safe; it does not by itself prevent or trigger retries.

import { createAction } from '@activepieces/pieces-framework';

export const createTask = createAction({
  name: 'create_task',
  displayName: 'Create Task',
  description: 'Create a task in a project',
  aiMetadata: {
    description:
      'Create a new task in a given project, with optional assignee, due date, and labels. ' +
      'Each call creates a new task, so it is not idempotent. ' +
      'Use Update Task to modify an existing task instead.',
    idempotent: false,
  },
  props: {
    /* ... */
  },
  run: async (context) => {
    /* ... */
  },
});

audience

Available on actions only (triggers have no audience: they always start flows, which both humans and agents build):

audience: 'human' | 'ai' | 'both'

Mark an action human when it only makes sense with the builder around it, for example the generic custom API call, or composite actions whose inputs assume a person picking from dropdowns. Mark an action ai for atomic operations added specifically for agents that would clutter the human piece selector.

Writing actions agents can use well

Agents work best with actions that behave like clean API calls:

• Atomic over composite. One action should map to one capability with explicit inputs. Agents compose multi-step work themselves, so a focused Create Task beats a Create Task and Notify Channel.
• Explicit inputs. Every behavior should be reachable through a documented prop; agents fill inputs from the property schema, not from a UI.
• Describe the output. Pair the action with an output schema so both the data selector and agents know the shape of what comes back.
• Disambiguate in aiMetadata.description. When a piece has several similar actions, say which one to use when: that sentence is often what decides which action the search returns.