Manus AI Review: What the Agent Does, Costs and Risks

Manus AI is a general-purpose AI agent. You describe a job, and it plans and finishes the work inside its own cloud computer, then hands back a research report, a slide deck, a live website, a video or a running automation. It is made by Butterfly Effect Pte. Ltd., a company in Singapore. Paid use starts with Pro at $20 a month (4,000 credits),^[1] and the Free plan runs only the lightest mode. Our verdict is simple. Manus is strong when you want finished deliverables rather than answers. It is hard to budget, because credits burn at a rate Manus itself cannot predict before a task starts.

What happens when you hand Manus AI a job

Every task gets its own isolated cloud virtual machine with internet access, a file system and a browser. The agent plans the work, opens sites, writes and runs code and assembles files, and the machine goes to sleep when nothing is happening. What comes out is concrete: research reports, decks exported as editable PPTX or PDF, websites and full-stack web apps with a database and login hosted on Manus, videos, browser games, mobile app test builds and scheduled automations. The current generation, Manus 2.0, launched in September 2026 on an in-house agent harness called Cascade, which adds specialized capabilities only when a job needs them.

The mode menu has four entries. Manus 2.0 Max is meant for complex tasks, Manus 2.0 for most work, and Manus 2.0 Lite for everyday tasks and small edits. Manus Flex, still in beta, lets you plug in your own model key from providers such as OpenAI, Anthropic or Google AI Studio, though images, video and hosting still draw on credits. Free accounts get Lite in Agent mode plus Chat mode; Pro unlocks all three 2.0 modes. Chat reads images and PDFs and searches the web. It is not free, though. Both modes use credits, and you can switch a conversation from Chat to Agent but never back.

The pieces that go beyond a chat window

Wide Research is the feature most buyers ask about. It splits a list job across parallel sub-agents, each a full Manus instance with its own machine, then merges the results into a table or report. It is paid only, starts automatically when the agent decides a task can be split, runs 20 sub-tasks at a time and caps each one at 50 credits. Manus says it is not meant for single deep dives, sequential work or lists under 10 items. Automations run on a schedule as often as every five minutes, or fire on triggers from Gmail, Slack, Shopify, GitHub, RSS or a webhook.

Manus Studio (the desktop app, formerly Manus Desktop) brings the agent onto macOS on Apple Silicon and Windows 10 and 11. Intel Macs and Linux are not supported. Its My Computer feature runs command-line tools against folders you authorize, and Remote Control lets you drive the desktop by voice from the iOS or Android app, as long as the computer is on, awake and running Manus Studio. The connector catalog listed 416 entries as of November 2026, from Gmail and GitHub to HubSpot and Notion. Our Manus AI playbook covers modes, prompts and Wide Research in depth.

Where Manus AI keeps your data and how you sign in

Manus says it stores data in the US and Singapore. Sign-in uses Google, Apple, Microsoft, email or passkeys.

Credits decide what Manus AI really costs

The annual cards show $17, $34 and $167 a month billed yearly, but the charge is $200, $400 and $2,000 a year. By our arithmetic, that saves $40, $80 and $400 a year against monthly billing, and plan credits cost $5 per 1,000 on every tier from $20 to $200 a month. The same $40 card offers 12,000, 16,000 or 20,000 credits a month for $60, $80 or $100 a month in total.

Add-on packs work out to $10 per 1,000 credits, twice the plan rate, and only higher plans can buy them: fixed packs are sold only on the highest personal plan, pay-as-you-go top-ups need a personal plan of $200 a month or more (the $167-a-month annual plan qualifies), and on Team, Owners and Super Admins buy them in Team settings. On the Pro plans from $20 to $100 a month there are no add-on credits, so running out means waiting for the monthly refresh or moving up the selector.

The rules around those credits matter more than the price list. Tokens, virtual machine time and third-party APIs all draw on the balance, and Manus cannot estimate a task's credit cost in advance; it tells users to treat any figure the agent offers as a hallucination. Monthly plan credits do not roll over, on annual plans either, and the 300 daily refresh credits work only in Manus 2.0 Lite, on every plan.^[2] Add-on credits, on the plans allowed to buy them, stay usable only while a paid plan is active.

Refunds are narrow. The Terms treat payments as final except where law requires, so consumers in the EU, the UK and Turkey keep a 14-day withdrawal right, and partial refunds happen only at Manus's discretion. Credits come back for verified platform faults after an AI review of the shared task link; Manus's help pages disagree on whether a poor result qualifies, so confirm with Manus in writing. Apple purchases are refunded by Apple. Website hosting and Cloud Computer are billed separately from the plan. Our Manus AI pricing guide walks through every tier, who can buy add-on credits and each separate bill.

Is Manus AI safe for company work?

The paperwork is solid. Manus's Trust Center lists SOC 2 Type II, SOC 2 Type I, ISO 27001 and ISO 27701, with the reports available to download, and its listed subprocessors (Google Cloud, Microsoft Azure AI Foundry, AWS and Cloudflare) all sit in the US. Tasks on individual accounts are private by default. Team accounts differ. There, the Owner and Super Admins can access every member's tasks, and the Owner can export that session data, so staff should know their work is visible.

Training needs careful reading. The Privacy Policy says Manus may build aggregated or de-identified data and use it to develop and improve its models, and users in some jurisdictions may opt out by email. Team plans include a data-training opt-out. You own your inputs and outputs and may use them commercially, though output may not be unique or protectable. Nobody under 18 may use the service.

The bigger exposure is what the agent can touch. In Manus Studio the default permission mode reviews safe actions automatically, yet even the approval modes can still read any file your user account can open, and Full access skips approval entirely. Through Browser Operator it acts inside your logged-in sessions, and the Privacy Policy says page content from the sites you authorize, clicks and form input go to Manus servers; only your passwords are not seen or stored. The Terms say agent actions may be attributed to you, and revoking access does not undo what is done. Manus also warns that output can sound confident and still be wrong, including invented citations. Our buyer's checklist for Manus AI turns these points into questions to settle before rollout.

How we score Manus AI

We weight capability and output range heavily for an agent like this, because the reason to pay is the finished work it returns. We treat cost predictability and support as real risks rather than footnotes, since both shape what a small team actually spends and how fast it recovers from a failed task. Sentiment counts, but as one axis among several. The full breakdown sits below.

That lands Manus AI at 4.0/5. Capability, AI and the range of connectors pull the number up; value for money and support pull it down. The public review platforms tell a split story that is worth reading by source rather than as one average.

We read Trustpilot as a complaint map, not a trend. The Manus profile there is unclaimed, and Manus has not invited reviews, so the page collects people who arrived angry. Their themes are consistent: support that relies on automated bots and makes a human hard to reach, disputed transactions, unexpected charges and credits that vanish fast. One reviewer reported losing a large unused balance after a mid-cycle upgrade. That is a single account, but it matches the official rule that unused monthly plan credits expire each cycle.

Where Manus AI pays off, and where it misfires

Manus pays off for owner-operators and small teams who want a finished artifact from one instruction. Think a market scan across dozens of companies, a deck built from a spreadsheet, a hosted landing page with a form, or a weekly report that assembles itself on schedule. The 20 tasks at once on every paid plan let one person run several jobs in parallel. Teams get pooled credits, free guest seats and an admin dashboard from $20 per seat a month.

It misfires when the budget is fixed to the dollar. A long task can drain a monthly allowance, and you learn the cost only afterwards in Settings. Free is a taste, not a workplace: by our arithmetic its 1,500-credit monthly cap equals five days of 300 refresh credits, all in Lite, with no Wide Research, video generation or custom domain. Skip it, too, if your desktop runs Linux or an Intel Mac. Start with a monthly Pro plan, watch Usage for one cycle, and only then consider yearly billing.

The bottom line

Manus AI is a genuine agent, not a chatbot with extra buttons: it gets a whole machine per task and returns work you can ship, from decks to hosted web apps. Our score reflects that range against two persistent weaknesses: costs you cannot see in advance and support that Trustpilot complaints describe as leaning on bots. Buy it month to month first, measure what your real tasks consume, and commit to a yearly plan only once the numbers hold.