I’ve spent the better part of a year building GloriaMundo, and today it’s live. I want to tell you what it is, why it exists, and what you can actually do with it right now.

The gap that bothered me

There are two kinds of automation tools, and neither one felt right.

On one side, you’ve got platforms like Zapier and Make. They’re reliable. They’re mature. But they require you to manually wire up every node, every connection, every conditional branch. You’re the architect and the builder. For a simple “when X happens, do Y,” that’s fine. For anything involving intelligence — summarising, classifying, generating, deciding — you hit a wall. These tools don’t think. They just plumb.

On the other side, you’ve got AI agents. Manus, OpenClaw, ChatGPT’s Operator. They’re impressive. You tell them what you want, and they figure it out. But here’s the catch: you can’t see what they’re going to do before they do it. You can’t schedule them. You can’t reuse them. You can’t set spending limits. And when something goes wrong at 3am, there’s no audit trail to tell you what happened.

The first group gives you control without intelligence. The second gives you intelligence without control.

GloriaMundo sits in the gap between them. You describe what you want in plain English. The AI builds a structured, inspectable workflow. You preview exactly what it will do — every step, every API call, every piece of generated content, every estimated cost — before anything executes. Then you approve it, run it, schedule it, and it works the same way every time.

That’s the idea. Let me show you what’s actually in the box.

How it works: describe, generate, preview, approve, run

The core loop is five steps.

Describe. You open the builder and type what you want in natural language. “Every morning at 9am, check my Gmail for invoices, extract the amounts, log them to Google Sheets, and send me a Slack summary.” Or something more complex: “When a new lead comes in via webhook, enrich their data, use AI to score them 1-100, then route high-scoring leads to the sales team with Slack alerts and HubSpot deals while low-scoring leads go into a nurture campaign.”

Generate. GloriaMundo’s Architect translates your description into a structured workflow with typed steps, execution strategies, and data flow between them. It uses dynamic tool discovery to find the right integrations from over 800 available apps. This isn’t a template lookup — it’s genuine AI reasoning about how to accomplish your goal, including which services to use, what data to pass between steps, and where to add conditional logic.

Preview. This is the part I’m most proud of. Before anything runs, you get a full Virtual Run — our Glass Box preview. READ operations (like fetching your emails or checking a spreadsheet) execute live against real data so you can see actual results. WRITE operations (like sending an email or posting to Slack) are simulated, with AI-generated preview content showing you exactly what would be sent. You see the estimated cost for the entire run. You see which integrations you’ve connected and which ones still need OAuth setup. You see every decision point and data transformation.

Approve. You review the preview. If something’s off, you modify it — either by chatting with the AI (“make the Slack message more concise” or “add an error notification if the API call fails”) or by editing the workflow directly in the visual editor. When it looks right, you approve.

Run. The workflow executes. If you set it up with a cron schedule, it runs on that schedule. If it’s webhook-triggered, it waits for the event. Every run is logged in full — inputs, outputs, costs, timing — in the action log.

What you can actually do with it

Let me walk through the features that exist right now, because I think specifics are more useful than adjectives.

Natural language workflow builder

You describe what you want, and the system builds it. The Architect understands service-specific concepts — it knows the difference between a Slack channel message and a direct message, between a Gmail send and a draft, between a Linear issue and a Linear project. It chooses execution strategies for each step and sets up data flow between them automatically.

For complex requests, the system can handle parallel branches (do these three things simultaneously), conditional logic (if the score is above 70, do X; otherwise do Y), and sub-agent spawning (delegate a subtask to a specialised agent).

The visual workflow editor

Every generated workflow appears on a node-based canvas. You can drag nodes, rewire connections, edit step parameters, and restructure the flow visually. This is a working editor, not a read-only display. You can add steps manually, remove steps, and adjust the data flowing between them.

The editor shows you the workflow as it actually exists: triggers at the top, steps flowing downward, parallel branches spreading horizontally, conditions forking into separate paths. What you see is what will execute.

Glass Box Virtual Run

I’ve described this above, but it’s worth emphasising what makes it different. Most preview systems show you a static “this is what will happen” description. Virtual Run actually executes the safe parts and simulates the rest with realistic generated content.

If your workflow reads from a Google Sheet and then sends a Slack message based on that data, the preview will show you the actual spreadsheet data and a realistic preview of the Slack message that would be generated from it. You see the real inputs flowing into simulated outputs.

The preview also runs connection readiness checks — it scans the workflow for required OAuth connections and tells you which ones are active, which need to be connected, and gives you direct links to set them up. No more “workflow failed because you forgot to connect Twitter.”

800+ integrations via Composio

Gmail, Slack, Google Sheets, HubSpot, Linear, Notion, Jira, GitHub, Twitter, LinkedIn, Salesforce, Airtable, Discord, Telegram — the list goes on. These aren’t shallow integrations. The system does dynamic schema introspection to understand each service’s parameters, handles OAuth connections, and resolves parameter naming inconsistencies automatically.

I’ll be honest: Composio’s parameter schemas aren’t always consistent between their SDK and MCP documentation, and some parameters are accepted in the schema but silently ignored by the backend. We’ve built a resolver that handles most of this automatically, but you may occasionally hit an edge case with newer integrations. We track known issues and fix them as they surface.

Step types

Workflows aren’t just API calls. A workflow can include:

Integration steps — actions against any of the 800+ connected services
AI steps — LLM-powered generation, analysis, summarisation, classification
Code execution — Python code running in an isolated E2B sandbox
Web search — research steps using Tavily or Serper with configurable depth
Browser automation — navigate, extract, and interact with web pages via BrightData
Conditional branches — if/else logic based on data from previous steps
Sub-agents — delegate complex subtasks to specialised agents
URL extraction — pull structured data from web pages
State management — read and write workflow state across steps

These compose naturally. A workflow might fetch data from an API, analyse it with an LLM, run a Python script to transform the output, then post results to Slack and log them to a spreadsheet — with a conditional branch that sends an alert if the analysis flags something urgent.

AI image generation

Workflows that involve social media or content creation can generate images as part of the pipeline. We support Google Vertex AI Imagen (both Fast and Ultra quality tiers), plus additional providers through OpenRouter including Gemini Image and FLUX. The system automatically selects appropriate aspect ratios based on the target platform — 16:9 for Twitter, 1:1 for Instagram, 1.91:1 for LinkedIn.

Scheduling and webhooks

Workflows can be triggered three ways: manually (you click Run), on a schedule (cron expressions with timezone support), or via webhook (an external event posts to your workflow’s endpoint). The scheduling system handles human-readable times (“every weekday at 9am London time”) and converts them to cron expressions automatically.

The skills library

Skills are reusable behaviours that your workflows can incorporate. They come in three flavours: prompt-based (custom system prompts that shape how AI steps work), code-based (Python functions that steps can call), and full workflow definitions (pre-built sequences you can drop into larger workflows). Think of them as building blocks you accumulate over time.

Demo workflows

When you first open the platform, we offer pre-built demo workflows that load directly onto your canvas. These are designed to show what’s possible: a sales lead qualification pipeline with enrichment, AI scoring, and branching; a content publishing automation; a customer support escalation system; a daily executive briefing. Each one has 13-20 nodes with parallel branches, conditional logic, and multiple AI steps. They’re real workflows, not mockups — you can preview and run them.

Spending controls

New accounts start with $20 in credits. The system uses a tiered spending limit that grows as you build payment history: $20, then $50, $100, $250, up to $1,000. You always see your credit balance in the builder. The Virtual Run shows estimated costs before you approve. If a workflow would exceed your remaining credits, the system tells you before it runs, not after.

Governance and safety

Every action passes through a governance gate before execution. This checks against a kill switch (for emergencies), blocked action lists, constitutional rules (immutable guardrails), policy evaluation, and approval requirements. Default policies require manual approval for financial transactions, public social media posts, and email sends. You can customise these — relax them for trusted workflows or tighten them for sensitive ones.

The audit log records every action, every step, every decision. If something goes wrong, you can trace exactly what happened, what data was passed, and why each decision was made.

The Iron Layer

Behind every workflow, a reliability layer handles the infrastructure you’d otherwise have to build yourself — automatic retries with backoff, circuit breakers, rate limit handling, and error recovery that resumes from where things left off. You don’t configure any of this. It’s built into every workflow by default.

If you’ve ever debugged retry logic at 2am, you’ll appreciate this. We’ll go deeper into the Iron Layer’s internals in a future engineering post.

Where things stand honestly

This is a soft launch. I’m one person, and while the platform works — the architecture is solid, the core loop functions, the integrations are real — there are rough edges.

The workflow editor is functional but the UX for complex branching could be smoother. Some Composio integrations behave slightly differently than their schemas suggest. The skills library works but doesn’t yet have a large catalogue of community-contributed skills. Browser automation is available but adds meaningful cost.

I’m not claiming this is finished. I’m claiming it’s useful, it works, and the fundamental approach — showing you everything before it runs — is the right one.

There’s a roadmap ahead. Workflow-as-API (publish any workflow as a secure endpoint your own apps can call) is designed and coming within a couple of months. Team collaboration features are in progress. The skill ecosystem will grow as people use it.

An invitation, not a pitch

If you’ve been frustrated by the gap between “powerful but dumb” automation tools and “smart but opaque” AI agents, GloriaMundo might be what you’ve been looking for.

I’m genuinely interested in feedback at this stage. What works, what doesn’t, what you wish it did. The platform is live at gloriamundo.com, and you can get started with $20 in free credits to try it properly.

If you’d rather just talk about the approach, I’m happy to do that too. The Glass Box philosophy — the idea that you should see what your AI will do before it does it — is something I feel strongly about, and I’d love to hear whether it resonates with how you think about AI automation.

Thanks for reading. Time to go build something.

Andy Surtees, Founder — GloriaMundo