Gumloop AI Automation Tutorial: Build Your First AI Flow in 30 Minutes

You open a competitor page, copy the pricing table into a spreadsheet by hand, paste the feature list into ChatGPT, wait for the summary, and rewrite it into a Slack message for your team. By the third page you are doing the same four steps again. That loop is exactly what Gumloop is built to remove. This tutorial builds one real flow that scrapes a page, asks an AI model to analyze it, and delivers the result, so you can judge whether Gumloop earns a place in your stack before you pay for it.

Quick answer

Create a free Gumloop account, build a flow with an Input node, a Web Scraper node, and an Ask AI node, connect them in order, run it once on a test URL, and schedule it only after a clean run.

The hard part is not connecting the nodes, it is keeping the AI output structured and the credit bill predictable, and both problems are decided by the prompt and the model tier you pick in the Ask AI node.

Checked on 2026-11-12 against the official Gumloop documentation and the Gumloop pricing page. Interfaces change fast, so if a menu or label differs, trust the linked page.

Evidence: what Gumloop is, and what it costs

Gumloop describes itself in its own documentation as an AI agent platform: “Build AI agents that use your tools and data to do real work, then automate them with triggers, schedules, and the API.” Gumloop puts an AI reasoning step in the middle of the flow, so the same flow can read an unstructured page, decide what matters, and write a structured answer. Its canvas mixes classic control flow with AI nodes, a browser automation engine, and Gummie, a built-in AI assistant.

The pricing model is credit based. Pro costs $37 per month and includes 20,000 credits, split as 7,400 base credits plus 12,600 bonus credits, with each credit indexed at $0.005. Gumloop bills tokens, compute, and paid tool calls at list price, then adds an 8 percent orchestration fee. Bring your own model API keys and token costs drop to zero, but the fee rises to 16 percent. Unused credits do not roll over, pay as you go bills at $0.005 per credit and is off by default, and every new account starts with a 14-day free Pro trial.

Prerequisites

• A Gumloop account. Sign up at gumloop.com; the 14-day free Pro trial covers everything in this tutorial.
• One public web page to analyze, such as a public docs page or a company homepage, not a page behind a login.
• A destination for the result: a Google Sheets account, or a Slack workspace if you prefer a message instead of a row.
• About 30 minutes, and a willingness to run the flow three or four times while you tune the prompt.

Step 1: Create your first flow

Sign in and open your personal workspace. Click Create Flow to open the canvas. Gumloop offers templates, and Gummie can draft a flow from a sentence, but build this one by hand so you learn where each piece lives. Name it something plain, such as “Competitor page analyzer”.

Step 2: Add an Input node for the target URL

Click the plus sign on the canvas and add an Input node. Add a text field named “Page URL” and paste a public URL into its default value, for example the Gumloop homepage at https://www.gumloop.com. Example, replace with your values: any public page you actually want analyzed works here.

Making the URL an input instead of hardcoding it inside the scraper is the difference between a demo and a reusable flow.

Step 3: Add the Web Scraper node

Add a Web Scraper node from the Core Nodes section and connect the Input node to it. Drag the Page URL field into the scraper’s URL field so the scraper reads the input at run time. Run the flow once now, before you add anything else.

If the run returns almost nothing, the page probably renders content with JavaScript after load, sits behind a bot check, or needs a login. Try a simpler public page first, such as a docs page. People waste the most time here by adding AI nodes on top of an empty scrape and then debugging the wrong node. A scraper that returns text is the only foundation worth building on.

Step 4: Add the Ask AI node, the brain of the flow

Add an Ask AI node and connect it to the Web Scraper output. Drag the scraped text field into the prompt box, then write a prompt that asks for a structured answer, for example: “Analyze this website text and extract the company’s value proposition, target audience, and pricing model, if mentioned. Return three short bullet lists with no extra commentary.”

Pick a model deliberately. For extraction and summarization, a small or mid-tier model is usually enough, and it costs a fraction of a flagship model per run. The DataCamp walkthrough of the same flow picks GPT-4.1 Mini for this exact step, a sensible default: cheap enough to run daily, capable enough to extract structured fields. Run the flow again and read the output like a skeptic. Does it quote the page, or invent a pricing tier the page never mentions? AI nodes hallucinate most often when the scraped text is thin, so a vague answer usually means the scrape in Step 3 was weak. Add one instruction, “If a field is not on the page, write ’not stated’ instead of guessing,” and run it again.

Step 5: Send the result somewhere useful

Connect the Ask AI output to a destination. A Google Sheets node appends one row per run, which builds a research log you can filter later. A Slack node posts the summary to a channel, which is better when a human needs to act on it the same day. Map the fields explicitly: page URL, value proposition, target audience, pricing model, and run timestamp. Without a timestamp, a row is useless in a month.

Step 6: Connect an MCP server for live tools

This is the step that turns a static flow into an agent. Gumloop supports MCP in two directions. First, give an Ask AI node extra tools: save the server under Credentials as an MCP Server, then open the Ask AI node, enable the Connect MCP Server option, and pick the saved credential. Second, Gumloop hosts its own MCP servers, such as the Firecrawl MCP server for reading the web, usable inside a flow or from Claude and Cursor with the same server URL and no API key of your own. The screens are covered in the MCP Server credential documentation and the MCP Client node documentation. We could not verify every menu label in the live canvas; vendor docs say the credential type is “MCP Server” under Add Credentials. If a label differs, trust the linked docs.

A practical first MCP test is to connect a search server and ask the flow to check whether the scraped page mentions a feature the vendor announced elsewhere. If the output cites a page it could only have reached through the tool, the connection works.

Step 7: Schedule it, then watch the first three runs

Add a schedule trigger only after the manual runs are clean. Start with a daily schedule, not hourly, and watch the first three scheduled runs. Scheduled runs fail in ways manual runs do not: a target site changes its layout overnight, a model provider has an outage, or the flow burns through the monthly credit grant faster than expected. Check the run history daily for the first week.

Copy-paste template: a reusable competitor watch flow

Use this as the starting shape for any research flow. Replace the bracketed values with your own.

FLOW: Competitor page analyzer
INPUT: Page URL (text) — Example — replace with your values: https://www.gumloop.com
NODES:
  1. Web Scraper — URL = {{Page URL}}
  2. Ask AI — small tier model (example: GPT-4.1 Mini)
     PROMPT: "Return: value proposition, target audience, pricing
     model (else 'not stated'), one change vs last run (else 'no change').
     Do not invent facts."
  3. Google Sheets — append row: URL, date, value proposition,
     target audience, pricing model, flagged change
TRIGGER: manual until 3 clean runs, then daily schedule

The “no change” line is the most valuable part. Without it, every run produces a fresh summary that looks new even when nothing changed, and the sheet fills with noise.

When not to use this / Limits

Do not build a Gumloop flow for exact numbers from a page that fights scraping, such as a dashboard behind a login or a site with aggressive bot protection. The browser engine helps, but it is not a guarantee, and a flow that silently returns partial data is worse than no flow. Do not use it for high-volume scraping: credit billing makes bulk extraction expensive fast. Do not use it where a wrong AI answer has real consequences, such as quoting a price to a customer, without a human review step in between. And do not use it as a system of record: the sheet or CRM it writes to is the record, not the flow. If you need self-hosted control, versioned workflows in git, or a fixed per-execution price instead of credit billing, n8n is the better fit, as our n8n vs Make comparison for AI agents explains.

Cost math: what one daily run actually costs

Use Gumloop’s own numbers. Pro is $37 a month for 20,000 credits, each indexed at $0.005. Suppose one run costs 40 credits: a small-model Ask AI call plus one scrape, before the fee. At one run a day for 30 days, that is 1,200 credits, about 6 percent of the monthly grant. At one run an hour, it is 28,800 credits, which blows past the 20,000 grant in about three weeks, and the flow stops unless you enable pay as you go. Add the 8 percent orchestration fee, or 16 percent with your own model keys. Run the flow five times, take the highest per-run cost, and multiply by the schedule.

Troubleshooting

Common mistakes

• Hardcoding the URL inside the scraper. The flow works once and cannot be reused. Put the URL in an Input node so the same flow handles a list of pages.
• Picking the strongest model for extraction. A flagship model costs several times more per run for work a small model does just as well. Downgrade first, upgrade only when output quality drops.
• Scheduling before three clean manual runs. A schedule multiplies every bug by every run. Manual runs are free debugging; scheduled runs are paid debugging.
• Trusting the summary without reading the scrape. When the AI invents a fact, the scrape is usually the weak link. Read the scraper output first.

Test before you rely on it

Run this checklist before the flow goes on a schedule:

1. Three manual runs in a row return complete, accurate output.
2. The scraper output contains real page text, not an error page or a cookie banner.
3. Every field is either filled from the page or marked “not stated”, never invented.
4. The destination row or message carries the run timestamp and the source URL.
5. The per-run credit cost is measured, and the monthly total at the planned schedule fits inside 20,000 credits.
6. A failed run is visible: you know where the alert goes.

FAQ

Is Gumloop free to try? Yes, a 14-day free Pro trial. After it, Pro is $37 per month with 20,000 credits, and unused credits do not roll over.

How is Gumloop different from n8n or Make? Gumloop is agent-first: AI reasoning nodes sit at the center of a flow. n8n and Make are workflow-first, with AI as one node type among many.

What does a Gumloop credit actually cost? One credit is indexed at $0.005. Pro includes 20,000 credits a month for $37, plus an 8 percent orchestration fee, rising to 16 percent with your own model API keys.

Can Gumloop call external tools through MCP? Yes. Save an MCP server as a credential, then enable it inside an Ask AI node or use the MCP Client node. Gumloop also hosts its own MCP servers, such as the Firecrawl MCP server.

Does the Web Scraper node handle JavaScript-heavy pages? It handles rendered pages better than a plain HTTP fetch, because the browser engine loads the page and extracts rendered text. Login walls and bot protection can still fail, so test the exact target URL first.

What happens when I run out of credits mid-month? Your agents stop and Gumloop prompts you to enable pay as you go at $0.005 per credit. Pay as you go is off by default.

Conclusion

Gumloop earns its place when a flow needs judgment, not just routing: read a page, decide what matters, and write a structured answer. The build itself takes half an hour. The discipline that makes it last is unglamorous: a structured prompt, a cheap model tier, a measured credit cost, and a schedule you only enable after clean manual runs. If this flow fails at 3 a.m., can your team see what happened? If not, add the timestamp, the run history check, and the credit alert first.

Next, compare costs in our n8n vs Make comparison for AI agents, or go deeper with our MCP server tutorial.

Frequently asked questions