# Gumloop Web Scraping: Browser, Connectors and Sandbox Setup

> How to scrape the web with Gumloop agents: Web Fetch, Firecrawl and Apify connectors, the cloud browser with logins and handoffs, and what each costs.

_Source: https://professionalstoolkit.com/articles/gumloop-web-scraping — The Professional's Toolkit · updated November 2026_

---

Gumloop has no single scraping button. You scrape by giving an agent the right tool for the page: the built-in Web Search and Web Fetch abilities for public pages, no-auth connectors such as Firecrawl, Apify, Parallel, Exa and Reducto for structured extraction, a cloud Chromium browser for sites behind a login, and a Python sandbox with scrapy and playwright for custom parsing. The key number is the browser's price. It has no separate charge and is billed as compute at **5 credits per minute** of active processing, plus model usage [1]. Our verdict: Gumloop suits recurring, sourced research that lands in a sheet or a dashboard. It does not promise to get past bot detection.


**5 credits a minute** — Agent browser compute, plus model usage; proxy, replays and handoff waiting are not billed.



- **5** — Credits per session-minute of active browser and compute time
- **20,000** — Chrome cookies per sign-in import into a browser profile
- **6 hours** — Longest background command in the code sandbox


## Step 1: choose the lightest tool that reaches the page

Order matters here, because each heavier tool costs more time and credits. Gumloop's docs recommend a connector first, then the site's API or a custom MCP server, and the agent browser last, for any site a person could use by hand. For public pages, the built-in Web Fetch ability reads a URL directly, and Web Search queries the web through a provider you choose: Exa, Parallel, Firecrawl or the model's own native search. Both abilities come with every agent. Start there.

One setup detail trips people up. An agent prefers only the connectors added to that agent, so a connector you linked in Settings alone does not count. If the browser is switched on and the connector sits only in Settings, the agent may open the website instead, which burns compute minutes for a job a single tool call could have done. Add the connector on the agent itself, then sign in. The same rule applies to every scraping connector in the next step, and it is the first thing to check when an agent behaves oddly.


> 💡 **Try a connector before the browser:** Gumloop's docs recommend a connector, then an API or custom MCP server, and the browser last. Add the connector on the agent itself, not only in Settings.




> ⚠ **AI-created triggers bill every poll:** A custom AI trigger costs at least 1 credit per poll even when nothing changes, so a frequent schedule drains the shared credit pool.

## Step 2: attach a scraping connector to the agent

Gumloop runs a group of connectors that work on the first run with no account linking, either because the data is public or because Gumloop supplies a managed API key. The research and web-data set is Exa, Parallel, Firecrawl, Apify and Reducto, each with an optional bring-your-own key. Separate connectors for YouTube, Instagram, TikTok and Google Maps scrape public data only and accept no key of your own. For company and SEO data, Apollo, Brandfetch and DataForSEO sit in the same no-auth group [2].

Two exceptions need your own account. Firecrawl's interactive browser tools and Apify's full-permission Actors work only with your own key. The Apify connector can also search the Apify Store for a ready scraper, run it and return the results, for example LinkedIn profiles, company pages or job postings (added May 2026). Firecrawl's scrape tool keeps a persistent profile, so a logged-in session carries across scrapes. **Managed-key calls are billed in Gumloop credits**, with each tool's price listed on its page in Gumloop's connector directory. With your own key, the call routes through your vendor account and drops to Gumloop's 1-credit minimum. The docs advise your own key for higher rate limits or for tools the managed key does not cover.

## Step 3: turn on the agent browser for logins and JavaScript

When no connector or API reaches the data, the agent browser takes over. It is real Chromium running on the agent's cloud computer, isolated from your own machine, and it handles JavaScript-heavy sites, sign-ins, forms, uploads and downloads. It went live for every agent in October 2026. New custom agents get it on by default when the creator's role allows; older agents stay off until someone turns it on. On Enterprise, the browser is a custom-role feature that starts off and is checked for whoever runs the agent, not for its owner.

### Logins, profiles and handoffs

Saved logins live encrypted in the Vault, or come from a 1Password service account with a read-only token, and passwords and 2FA keys never reach the agent. Every fill is checked against the login's site first. Authenticator-app codes are generated automatically; SMS or email codes need a person. On a login, 2FA code, CAPTCHA or purchase confirmation, the agent pauses and hands the browser to the person whose message it is working on, who can answer in the live browser, from Slack buttons or, on Enterprise, where Gumball can text you, from a secure link sent by text. **Waiting is free.** Sign-ins save to a browser profile at the end of each turn, and incognito chats, an Enterprise feature, never save or restore them.

You can import existing Chrome sign-ins through Gumloop's Chrome extension, a macOS terminal script or the CLI, up to 20,000 cookies per import. Gumloop's Chrome extension, listed as an early release, records actions you take in the browser and replays them as a repeatable automation, and it can run Gumloop automations on the page you are viewing. The docs also warn that the password protections defend against a web page steering the agent, not against an agent set on extracting a value from its own sandbox. Bind only the logins a job needs.

### Proxy, mobile mode and blocked sites

The browser routes through a residential proxy at no extra charge, set to Off, Auto or a chosen country; changing it restarts the browser but keeps sign-ins. Mobile mode emulates an Android phone for testing websites, not native apps. Where a site exposes its own WebMCP tools, the agent calls them instead of clicking. Gumloop describes stealth browsing and proxied connections only as ways to reduce interruptions on sites that flag automation. Some sites block every automated browser, and for those the docs point you back to a connector, Web Fetch for public pages, or the site's API with a stored secret [3].


> ⚠ **Some sites block every bot:** Gumloop's stealth browsing and proxy only reduce interruptions on sites that flag automation. For sites that block all automated browsers, use a connector, Web Fetch or the site's API.


## Step 4: parse in the code sandbox when the page is messy

Every agent has a code sandbox switched on with no configuration, and each conversation runs in its own isolated cloud VM with full internet access. It ships the usual web libraries, including requests, beautifulsoup4, scrapy, selenium and playwright, so an agent can write and run its own scraping script when a connector returns raw HTML that needs cleaning. Limits are concrete: 30 minutes for a command the agent waits on, up to 6 hours for one in the background, and up to 300 MB per file. The sandbox runs headless only.

Packages installed with pip or npm last for the current chat only. To keep one in every chat, list it in `.workspace/agent/requirements.txt`, which only members with edit access can change. API keys for a target site belong in the Vault as agent secrets, handed to code as environment variables; Gumloop says this is not a guarantee that code cannot expose a credential. Sandbox code can also call connector tools through the preinstalled gumloop Python SDK, under the same tool permissions and approvals as the agent. That lets one script fetch with Firecrawl and write rows to a sheet.

## Step 5: start from the ready-made scraping agent and schedule it

The fastest start is Gumloop's AI web scraping agent in its use-case gallery. Clicking Use this agent creates a preconfigured copy in your workspace, and you authorize each tool with your own credentials. It pairs Parallel as the reasoning engine with Firecrawl as the scraping engine and the Python sandbox. It can pull a single page, map a whole site, crawl a section, render JavaScript, and click, fill or load more before reading. Results come back as a CSV, a spreadsheet, markdown or a dashboard, with a source URL behind every fact and a confidence read on uncertain fields [4].

For prospect lists, the LinkedIn agent connects Apify, Parallel, Google Sheets and Google Docs. It uses public data only and never logs into a LinkedIn account, so it sends no connection requests, messages or posts, and it leaves a field blank rather than inventing an email. Our [guide for sales teams](/articles/gumloop-for-sales-teams) shows where it fits a pipeline. Once a browser run works, Gumloop suggests saving the steps as a skill, so the next run follows a procedure instead of working it out again.

Recurring scrapes need a trigger. A scheduled trigger reruns the job on a plain-language schedule. The Parallel Web Monitor trigger needs no credentials and watches the whole web for changes matching a natural-language query, hourly, daily or weekly. AI-created custom triggers are the costly option, because they bill every poll whether or not anything fires. Our [AI playbook](/articles/gumloop-ai-playbook) covers triggers, approvals and Reflections in depth.


> 💡 **Save a working run as a skill:** Once the browser completes a scrape, save the steps as a skill so later runs follow the procedure instead of rediscovering it.


## What a scrape draws from your credits

| Method | What it is billed as |
|---|---|
| Connector call (Exa, Firecrawl, Apify...) | At least 1 credit per successful call, plus the tool's own charge on a managed key; failed calls free |
| Agent browser | Compute at 5 credits a minute plus model usage; proxy, replays, live view and handoff waiting free |
| Code sandbox | Compute for running time, no per-call charge |
| AI-created trigger | Every poll, at least 1 credit, plus any decision-model call |

Each method spends credits in a different place, and the chat details drawer shows the split after every run. By our arithmetic, browser compute alone comes to 300 credits, or $1.50, per hour of active processing, which means 200 session-minutes per 1,000 credits before the 8% orchestration fee. An AI-created trigger left at its default of every 5 minutes, the shortest interval, spends at least 288 credits a day, about 8,640 in 30 days, before any decision-model calls. Gumloop Pro is $37 a month for the whole organization's shared pool of 20,000 credits, and the same agent can spend 2 credits on one run and 200 on another; our [pricing breakdown](/articles/gumloop-pricing) covers overage, the cap and the monthly reset.


## FAQ

**Can several teammates share one logged-in browser profile?**

Only on Enterprise: an admin enables the Agent-owned credentials feature on a custom role, and the agent can then use one shared agent-owned profile, though not while it is shared with Anyone. On Pro, profiles are user-owned, so each person signs in once with their own account.

**How many 1Password vaults can one Gumloop agent read?**

An agent can use all vaults the read-only service account sees or up to 10 chosen vaults, with up to 5,000 logins per turn; Private vaults are not visible to it.

**Is the agent browser on by default for Enterprise teams?**

No. On Enterprise organizations the Agent browser feature is off by default in custom roles, and an admin must enable it per role; a teammate whose role lacks it gets the agent without the browser.

**How many cookies can a Gumloop browser profile hold?**

A profile holds up to 25,000 cookies or 24 MB, and cookie import brings cookies only, not local storage.


## Where Gumloop scraping earns its place

Gumloop is a strong fit when the scrape is a step in a larger job: research that needs sources, a lead list that must land in a sheet, a competitor page watched every week, or a portal that needs a login and a human for the 2FA code. The connector-first order keeps most runs cheap, and the browser covers what connectors cannot reach. It is a weaker fit for one-off bulk crawls of hostile sites. Some sites block every automated browser, and Gumloop sells no way around that, so check your target first.



## Sources

[1] Gumloop documentation: Agent browser — https://docs.gumloop.com/core-concepts/agent_browser (2026-11)
[2] Gumloop documentation: No-auth connectors — https://docs.gumloop.com/nodes/mcp/no_auth_connectors (2026-11)
[3] Gumloop documentation: What if a website blocks my agent's browser? — https://docs.gumloop.com/help/browser/website-blocks-agent-browser (2026-11)
[4] AI web scraping agent — https://www.gumloop.com/use-cases/ai-web-scraping-agent (2026-11)
[5] Gumloop pricing — https://www.gumloop.com/pricing
[6] Gumloop documentation: Credits — https://docs.gumloop.com/core-concepts/credits
[7] Gumloop Terms of Service — https://www.gumloop.com/tos
[8] G2: Gumloop reviews — https://www.g2.com/products/gumloop/reviews