Step Types¶

Workflows are built from steps, each of which performs a specific action. GloriaMundo supports the following step types.

Integration¶

Connect to hundreds of external apps.

Integration steps use Composio to interact with external services like Gmail, Slack, Google Sheets, HubSpot, and more. Each integration step executes a specific action (e.g., send an email, create a contact, fetch spreadsheet data).

• Requires an OAuth or API key connection to the target app
• Parameters are mapped automatically from your prompt
• Supports template variables: {{trigger.field_name}} or {{step_id.field}}

Example Actions¶

• GMAIL_SEND_EMAIL — Send an email via Gmail
• SLACK_SEND_MESSAGE — Post a message to a Slack channel
• GOOGLESHEETS_READ_ROWS — Read data from a Google Sheet
• HUBSPOT_CREATE_CONTACT — Create a new contact in HubSpot

LLM (AI Reasoning)¶

Use AI models for reasoning, generation, and transformation.

LLM steps send a prompt to a language model and return the response. Use them for:

• Summarising content
• Drafting emails or messages
• Analysing sentiment
• Extracting structured data from text
• Making decisions based on context

Configuration¶

• Prompt — The instruction for the AI model
• Model — Optionally override the model used for this step (defaults to your preset's writer model)
• Parameters — Temperature, max tokens, and other model settings

Code¶

Run custom Python in a secure sandbox.

Code steps execute Python code in an isolated E2B sandbox environment. Use them for data transformation, calculations, or any logic that's easier to express as code.

Key Details¶

• Code runs in an isolated environment — it cannot access your local machine
• For single-input steps, access previous step data via the input_data variable
• For multi-input steps, use input_1, input_2, etc.
• The step output is the return value of the code (or the last expression)
• Test your code directly in the editor with mock values using the Test button

Example¶

# Filter emails from the last 24 hours
from datetime import datetime, timedelta

cutoff = datetime.now() - timedelta(hours=24)
recent = [e for e in input_data if e['date'] > cutoff.isoformat()]
{"count": len(recent), "emails": recent}

Conditional¶

Branch your workflow based on conditions.

Conditional steps evaluate a condition and route execution accordingly. Use them for:

• Filtering data (only proceed if certain criteria are met)
• Branching logic (do X if true, Y if false)
• Error handling (check if a previous step succeeded)

Configuration¶

• Condition — A string expression evaluated against available step data
• Parameters — Values to compare against

Agent (Sub-Agent)¶

Spawn a child workflow for complex sub-tasks.

Agent steps create a sub-workflow that runs independently within the parent workflow. Use them for:

• Tasks that need focused, multi-step reasoning
• Generating multiple perspectives on a topic
• Deep exploration of a subject

Configuration¶

• Objective — What the sub-agent should accomplish
• Budget limit — Maximum spend in dollars for this sub-agent (default $1.00, up to $100)
• Allowed integrations — Optionally restrict which apps the sub-agent can use
• Briefing — Context, constraints, and role/expertise information
• Parallelisation — Run multiple sub-agents in parallel with different inputs or perspectives, then aggregate results

Sub-agent costs are constrained by the configured budget limit and capped by the parent workflow's budget.

Browser Automation¶

Automate web interactions when no API integration exists.

Browser steps use headless browsers to navigate websites, click elements, fill forms, and extract content. Use them as a last resort when the app doesn't have a direct integration.

Available Actions¶

Tips¶

• Use specific CSS selectors to target elements reliably
• Add wait steps before interacting with dynamically loaded content
• Test with Virtual Run first to verify element selection

Web Search¶

Search the public web for information.

Web search steps query the web via Tavily or Serper and return structured results. Useful for:

• Researching companies, people, or topics
• Finding current information
• Gathering competitive intelligence

Configuration¶

• Query — The search query
• Site filter — Restrict results to a specific domain
• Time filter — Limit results to a time range
• Search depth — Standard (up to 10 results) or Deep (up to 20 results, higher cost)
• Search tool — Optionally force a specific provider (tavily or serper); if omitted, the platform auto-routes

Twitter/X Search¶

Search Twitter/X with native access.

Twitter search steps use Grok for direct Twitter/X data access, providing better coverage than web scraping.

Configuration¶

• Query — The search query
• Limit — Maximum number of results

URL Extract¶

Extract content from a specific URL.

URL extract steps fetch a web page and extract its content in a structured format.

Configuration¶

• URL — The page to fetch
• Selector — Optional CSS selector to extract a specific element
• Render JS — Whether to render JavaScript before extracting (slower but handles dynamic content)
• Output format — Text, HTML, Markdown, or JSON

State (Get/Set)¶

Persist values across workflow runs.

State steps allow you to save and retrieve values that persist between workflow executions. Use them for:

• Tracking counters or timestamps across runs
• Storing the last processed item to avoid duplicates
• Maintaining state for scheduled workflows

Configuration¶

• Key — A unique identifier for the stored value
• Value (set only) — The value to store
• Default (get only) — Value to return if the key doesn't exist
• TTL (set only) — How long to keep the value (in days)

Vision¶

Understand images and visual documents.

Vision steps describe, classify, or answer questions about an image. Use them for photographs, screenshots, diagrams, receipts, and other inputs where layout or visual context matters.

OCR¶

Extract text from images and PDFs.

OCR steps read screenshots, scanned pages, and PDF documents. Text-based PDFs use their embedded text where possible, up to the first 100 pages. Scanned or image-only PDFs are rendered and read page by page, up to the first 10 pages.

Structured Extract¶

Turn unstructured content into predictable fields.

Structured extraction steps apply a declared schema to text or document content. Later steps can reference the resulting fields without parsing free-form prose.

File Output¶

Create a file from workflow results.

File output steps create text, Markdown, CSV, HTML, or JSON files in Google Drive. They can optionally email the output through Gmail, and filenames and content can reference earlier step results.

Foreach¶

Repeat a group of steps for every item in a list.

Foreach steps resolve a list while the workflow runs, then execute a contained sub-workflow for each item. Concurrency and maximum-item controls keep the fan-out bounded.

The contained sub-workflow cannot include sub-agent, state read, state write, or nested foreach steps. Its steps also cannot use the notify_human error policy; use retry, skip, or abort, or move the human-approval step outside the foreach.

Platform Data Query¶

Query a dataset maintained by GloriaMundo.

Platform data query steps filter, sort, and retrieve rows from an available central datastore. These datasets are read-only to workflows and are available only when GloriaMundo has published a relevant datastore.