Gumloop Review: What It Is, What It Costs and Who It Fits

Gumloop, answered in one paragraph

Gumloop is a hosted platform where a team describes AI agents in plain language and lets them work across its apps: Gmail, Slack, Salesforce, Notion and 300+ other built-in connectors, plus any MCP server you add. One Pro plan costs $37 a month for the whole organization's shared pool of 20,000 credits, with unlimited agents, seats and teams ^[1]. There is no free plan. Our verdict is clear. It is a strong pick for business teams that want working agents inside Slack, Teams and email without running a server, and a weak one if you need a chatbot for customers or a fixed monthly bill.

Gumloop describes itself as a multiplayer agent builder: anyone on the team builds agents with any model and any integration, while IT keeps control of access and spend. That pitch matches the product's shape. An agent here is a set of instructions, a chosen model, the connectors it may touch, skills, company knowledge and triggers that start it without anyone typing. Gumloop is made by AgentHub Inc. Everything runs in Gumloop's cloud; the nearest thing to running it yourself is the Enterprise option to deploy into your own AWS, Azure or Google Cloud account, which Gumloop still operates.

The pieces a team actually builds with

Gumball, custom agents and the Legacy builder

Every account gets Gumball, a personal agent that needs no setup. It uses your own connectors, writes a Daily Chew report of what you missed, prepares meeting briefs, and drafts email replies in Smart Inbox without ever sending them. Ask it for a custom agent and it builds one after you approve a card describing the change. Custom agents are where team work happens. You pick one of 35+ models from Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot, MiniMax, Z.ai and Meta, or let Auto route each message to a lower-cost model that can handle it. Skills load only when relevant, and up to 10 subagents can work on one task at once.

Gumloop's older step-by-step builder, Workflows (once called Flows), still appears on the pricing page marked Legacy; Gumloop's current docs and product are built around agents. Agents can call workflows, Gumloop says. The Python SDK's legacy client, which only runs saved flows, already warns that it is deprecated. If you arrive from an old tutorial full of step diagrams, treat it as history and start from the agent editor or from a conversation with Gumball instead.

Triggers, Brain and a browser of its own

Triggers turn a chat assistant into a worker. An agent can run on a schedule as often as every minute, react to events in Gmail, Slack, Salesforce, HubSpot, Zendesk, Jira and other apps, accept a webhook, or follow a custom trigger the AI writes for itself. Company Brain indexes Drive, Notion, Slack, Confluence^[p1], SharePoint and other sources, then answers with citations; it has no OCR, so text inside scans is skipped. Since October 2026 every agent can also get a cloud Chromium browser that signs in, fills forms and downloads files. On a login, a 2FA code or a CAPTCHA, it hands control to a person.

Files the agents produce become versioned artifacts with their own links: slide decks, PDFs, spreadsheets, documents and interactive HTML pages. Live artifacts go further and rerun their scripts each time someone opens them, using the viewer's own access and credits. Code runs in an isolated cloud machine with Python and shell on every agent. Together these are why Gumloop reads less like an automation tool and more like a workspace of digital colleagues. The catch is that every one of those surfaces draws on the same credit pool.

Where people talk to the agents

Agents meet people where they already work. The standard Slack bot answers in threads, one agent per channel; a Custom Slack App, which needs a Slack workspace admin and about 10 to 15 minutes, adds direct messages and your own name and avatar. Microsoft Teams arrived in June 2026 with tighter limits: public standard channels only, an @mention in every message and no custom branding. Each agent can also have its own gumloop.ai email address, a hosted chat page, a voice call in the browser, an API endpoint, and an MCP connection that lets Claude, Cursor or ChatGPT use it. One agent can serve several channels at once.

What $37 a month really buys

Pro is the only self-serve plan, and Gumloop's pricing page says it starts at $37 a month. Gumloop's pricing page shows no higher Pro option; the only way to spend more on Pro is overage. By our arithmetic, $37 a month buys 7,400 credits at the $0.005 list price, so 12,600 of the 20,000 monthly credits are a bonus worth $63 at list price. Gumloop's pricing page prints only a monthly Pro price; it shows no annual plan. Enterprise is custom pricing through sales. The pool is shared. Every member, every Slack mention and every API call draws from the same 20,000 credits.

The harder question is what a credit buys, and Gumloop answers honestly: it varies. The same agent might spend 2 credits on a quick question and 200 on a deep research task, in Gumloop's own two examples ^[2]. Each run adds up model tokens at the provider's list price, at least 1 credit per successful connector call plus any charge the tool adds, 5 credits per minute of active compute, and an 8% orchestration fee on top. Gumloop's own docs example comes to 459 credits, about $2.30 by our arithmetic. Failed connector calls and time spent waiting for you cost nothing.

Four rules shape the monthly reality. Credits reset each month and do not roll over except on Enterprise. Overage is off by default, so agents and triggers simply stop when the pool runs dry; switched on, it costs $0.005 per credit and Pro always caps it, by default at 1,000,000 credits a billing period. The trial is one-time and lasts 14 days. It requires a card and becomes paid Pro unless you cancel before it ends . Under Gumloop's Terms, paid subscription fees are non-refundable except where the law requires otherwise. Our Gumloop pricing guide works through every cost line.

Where Enterprise begins, and why it matters on day one

The gate that bites first is not security but budgeting. On Pro, everyone shares one pool and no member can be given a personal cap. A single busy teammate or a chatty trigger drains the month for all. AI-created triggers poll every 5 minutes by default, and by our arithmetic one left at that interval uses at least 288 credits a day, about 8,640 credits in 30 days, before any model calls it makes. Admins can see usage across members on the Usage and Limits page, and alert emails arrive at 75% and 90% by default. That helps you notice a problem, not prevent one.

Data handling also splits by plan. On Pro and Pro trials, Gumloop's Terms and Privacy Policy let it use your inputs and outputs to develop and train its own AI models; on Enterprise it does not, unless the customer asks for it in writing . OpenAI and Anthropic are contractually committed not to train on data sent through Gumloop. Users outside the US have their data processed in the United States. Gumloop says it holds SOC 2 Type II, meets GDPR and covers HIPAA, with business associate agreements offered for eligible plans.

The teams it suits, and the jobs it does not

Gumloop fits marketing, sales and operations teams whose work already lives in Slack, Gmail, a CRM and shared drives. Its use-case gallery lists 21 marketing agents and 9 sales agents, among others, and clicking Use this agent creates a preconfigured copy in your workspace. You then sign each tool in with your own credentials, so the agent sees only what you can. Approvals are flexible: each connector can always run, ask each time, ask only before writes and deletes, or never run. For a team that wants agents doing real work without anyone managing servers, that combination is hard to beat.

It is the wrong tool for a public-facing assistant. Visitors to a hosted page must sign in with a Gumloop account, custom domains are not supported, and iframe embedding is not officially supported. An agent's email inbox answers only senders whose address matches a Gumloop account with access. A front end for your customers therefore means building on the API. Developers on Windows should know the Gumloop CLI runs on macOS and Linux only; the alternative is WSL or the Python SDK. And buyers who need a fixed cost per month will struggle. Credits move with every run.

How we score Gumloop

We weigh Gumloop on eight axes, each on its own evidence, with capability and cost transparency counting heavily for an agent platform billed on usage. Public reviews of Gumloop are still few on every platform, so the ratings below are early signals, not a trend. We read them as a check on our own reading of the docs, one axis among eight and never the verdict. The full breakdown follows.

Our composite lands at 4.0/5. Breadth carries it: agents, a personal assistant, triggers, knowledge, a browser and real channels in one plan priced for the whole organization. Two things hold it back. Costs per run are hard to forecast, and the controls that tame a shared pool sit behind Enterprise. If either rules it out, our Gumloop alternatives guide compares the other agent platforms, and the Gumloop buyer's checklist lists what to verify before the trial converts.

Is Gumloop worth putting a card down for?

Yes, if your team can name two or three recurring jobs that cross Slack, email and a CRM, and someone will watch the credit pool during the first month. The trial needs a card and turns into paid Pro unless you cancel before it ends. Use the 14 days to run those jobs through the chat drawer's cost lines, set approvals before any agent writes, and decide whether one shared pool is enough. If you need per-member caps, single sign-on or a customer-facing bot, the answer depends on Enterprise or another product. Our score reflects that split: excellent breadth for teams, real uncertainty for budgets.