In the spring of 2025 a short demo video from a previously unknown Chinese start-up broke through the usual AI news cycle. The product was called Manus, and the pitch was simple: give it a sentence, walk away, come back to a finished piece of work. Spreadsheets filled. Research reports written and cited. Travel booked. Slide decks drafted. All of it produced by an agent that opened its own browser, wrote its own files and planned its own next steps — the first thing in public that actually behaved the way the phrase “AI agent” had always implied.

The hype was enormous. Invitation codes traded on Discord servers for hundreds of pounds. The company’s servers fell over within hours of launch. A year and a half later, the dust has settled enough that we can give it a proper look. This is a UK operator’s guide to Manus AI in late 2026 — what it actually is, who built it, what it can honestly do for a small business here, and the one question most reviews quietly skip: whether you can run client data through it without breaching GDPR. If you want the shorter comparison, our Claude vs Manus for business head-to-head sits alongside this piece.

A note on the position I write from. Claudable trains UK small businesses on Claude Cowork and Claude’s agent features — computer use, Claude for Chrome, and the automation features built on them. That is a direct commercial conflict with Manus. I’ve tried to be honest where Manus is genuinely impressive, and honest where I think it isn’t ready for UK client work. You should read the compliance section in particular with an auditor’s eye rather than mine.

What Manus AI actually is

Manus is a general-purpose autonomous agent built by Butterfly Effect, a Chinese AI start-up founded by Xiao Hong and backed principally by Tencent and a handful of Chinese venture funds. The company was not a household name before Manus launched — it had spent its first two years building a less-noticed consumer product called Monica — and the sudden pivot to a flagship agent product caught the market by surprise.

The launch itself, in March 2025, was unusually dramatic. The product was announced with a short demo video and an invitation-only waitlist. Within 24 hours the waitlist ran to several hundred thousand people, invitation codes were being resold online, and most of the Chinese and English-language AI press was describing it — in the start-up’s own words — as “the world’s first general AI agent.” Whether that claim was accurate is a different question (OpenAI’s Operator, Anthropic’s computer-use Claude and Google’s Project Mariner were all live in some form at the time), but the branding stuck.

The product itself is a cloud-based agent that runs in what the company calls a “virtual machine sandbox.” Under the hood that is a Linux container with a browser, a file system, a shell, and a toolbox of helpers for things like PDF parsing, slide generation and spreadsheet editing. The agent plans its own steps, operates the browser the way a human would, saves working files, and — this is the part most other agents don’t do nearly as well — keeps running asynchronously after you close the tab. You can brief it, go to lunch, and come back to a finished draft an hour later.

Technically, Manus is not a single model. It is an orchestration layer sitting on top of foundation models from other providers — versions of Anthropic’s Claude and Alibaba’s Qwen have both been identified in its stack at different points. The clever part is not the underlying intelligence, which is rented. It is the planning loop, the tool harness, and the asynchronous execution environment. That is a real engineering achievement, and it is worth understanding on its own merits regardless of whether you ever use the product.

How Manus works, in practice

Using Manus feels unlike using Claude or ChatGPT. You don’t converse with it turn by turn. You brief it — usually in a paragraph or two — and then watch a split-screen: on the left, the agent’s running commentary of what it is doing; on the right, a live view of the browser and files it is operating. A typical session for a market-research task looks roughly like this.

The whole run is persistent. Close your laptop, come back in two hours, and the session is still there. That asynchronous quality is the single feature that most differentiates Manus from the interactive agents built by OpenAI, Google and Anthropic — those tend to run while you watch and fail quietly when you don’t.

Access, pricing and availability in the UK

For most of 2025 Manus was invitation-only. By mid-2026 the waitlist has effectively disappeared, and UK users can sign up directly with an email address. There is no UK-specific portal and no UK or EU data-residency option — you are using the same global service as everyone else, and the infrastructure sits principally in the United States and mainland China.

Pricing is credit-based, which is more complicated than a straight subscription. At the time of writing the published tiers are a Starter plan at roughly $19 per month for around 1,900 credits, a Pro plan at $39 per month for around 3,900 credits, and higher team tiers that add seats and credit pools. A simple research task will typically consume 200 to 400 credits. A heavier multi-step task — a full competitive teardown, say, or a travel booking that involves comparing options across half a dozen sites — will burn 600 to 1,000. In practice a daily user on the Starter tier runs out of credits in the first week of the month, and there is a meaningful step up in cost for anyone who wants to use the agent as part of a regular workflow.

There is no invoice-friendly UK billing entity. Payment is in dollars via Stripe to Butterfly Effect’s payment processor, which is fine for a sole trader but awkward for a mid-sized firm that wants to put the spend through accounts payable with a UK VAT invoice.

Who it’s for in a UK SMB

Manus is not a mass-market tool and it is not trying to be one. The honest audience in a UK small business looks like this.

It is not, in our view, for client-facing production work in regulated industries. Not for solicitors handling matter files. Not for accountancies handling client books. Not for recruiters handling candidate data. Not for anyone in healthcare, financial services or anything near the Data Protection Act’s special-category territory. The compliance section below explains why.

Ten realistic use cases

These are the tasks where Manus is genuinely useful today, with the honest caveats attached.

1. Market research with cited sources

Give it a market, a question and a geography, and it will return a structured report with sections, figures and inline links to every source. It is noticeably better than hand-crafted research for breadth and noticeably worse for judgement. Treat the output as a well-organised reading list with a draft summary on top.

2. Competitive teardowns

Point it at five competitors, specify what you want compared — pricing, positioning, messaging, feature set, reviews — and it will produce a side-by-side matrix. The quality of the matrix is directly proportional to how specific you are in the brief. Vague briefs produce vague tables.

3. Structured data extraction from public sites

Ask for a spreadsheet of UK B Corp certified agencies in Yorkshire with their LinkedIn URL and headcount, and it will do it. Row quality varies — expect to manually fix roughly one in six — but the saving against doing it by hand is enormous.

4. Drafting blog posts, then publishing them

With a brief, a tone example and credentials, Manus can write a long-form draft and push it to WordPress or Ghost. The writing quality is decent but generic; the useful mode is “draft so a human editor can finish in half the time,” not “publish without reading.”

5. Lead list enrichment

Upload a CSV of company names and it will go away and add industry, headcount, headquarters and recent news. Useful and risky in equal measure: if any of those names map to identifiable individuals, you have just exported personal data to a Chinese cloud (more on this below).

6. Travel booking

The flagship demo use case. In practice it works best for research — comparing flights, hotels and routes — and less well for the actual booking step, which still trips over CAPTCHAs, two-factor codes and card verifications roughly one time in three.

7. First-draft slide decks

Hand Manus a research document and ask for a ten-slide board update and it will produce something that looks presentable and is roughly 70% right. It is a genuine accelerator for anyone who regularly turns documents into decks; it is not a substitute for a human presenter’s judgement about what matters.

8. Expense categorisation and reconciliation

Upload a CSV of transactions and Manus will tag them, flag anomalies and produce a reconciliation note. Fine for internal, anonymous experiments. Not fine for live client books — both because of the data-residency problem and because the error rate, while low, is not low enough to avoid a human check.

9. SEO audits on public pages

This is a lovely fit. The target site is public, the auditing criteria are well-defined, the output is naturally structured. Manus will produce a readable audit of page titles, meta descriptions, internal linking, heading hierarchy and Core Web Vitals in minutes.

10. PR and brand monitoring

Ask it to sweep news, forums and social for mentions of your brand or a client’s brand over the last 30 days, categorise the sentiment and summarise what mattered. It is competitive with paid monitoring tools for a single-shot pass; it is not a replacement for a continuous monitoring platform.

Reliability: demo versus reality

The demos you have seen online are real but curated. Our experience, and the pattern in independent reviews, is that Manus completes roughly two-thirds of prepared tasks cleanly on the first attempt, roughly a quarter get partway there and need a human nudge, and roughly one in ten fails in ways that are hard to debug — a browser session that refuses to log in, a file write that silently corrupts, a plan that loops. Those numbers are improving steadily, but the honest framing for a 2026 adoption decision is: expect every output to need a human pass before it leaves your organisation.

Two specific failure modes are worth calling out. First, Manus handles English-language, Western-facing web content better than it handles UK-specific sites that lean heavily on Cloudflare bot protection. HMRC, several UK banks and the big UK job boards will frustrate it. Second, the agent is more confident than it ought to be about whether a task has succeeded. It will cheerfully tell you that it filed your quarterly return when what it actually did was fill in half a form and leave the page.

The UK GDPR problem

This is the section most reviews skip, and it is the single most important part of this guide for any UK business that handles client data.

Butterfly Effect is a Chinese company. Manus’s cloud infrastructure, per the company’s own documentation, is operated principally from the United States with elements in mainland China. There is no UK, EU or EEA region, and no public Standard Contractual Clauses — the Article 46 transfer mechanism UK law requires for exporting personal data to a country without an adequacy decision. Neither the United States nor China has an adequacy decision that extends to Manus’s setup (the EU’s Data Privacy Framework applies to certified US recipients only, and no equivalent exists for China).

What that means in practice for a UK data controller:

There are legitimate uses that do not trigger these issues — Manus running against entirely public data, not logged in to anything personal, with no uploaded files that contain personal data, is fine. The trouble is that the moment you connect it to a Gmail, a CRM or a client dataset to unlock its most useful workflows, you have crossed the line. For most UK businesses I speak to, that line is not worth crossing until Butterfly Effect offers UK or EU data residency, publishes SCCs and signs a UK-law data-processing agreement. At the time of writing, none of those exist.

When to use Claude’s agent features instead

Anthropic has been quieter in its agent launches than Butterfly Effect, and as a result Claude’s agent capabilities are less talked about than they ought to be. There are two in particular that cover a very large share of what people want Manus for.

Claude’s computer use gives the model control of a virtual desktop — mouse, keyboard, screen — so it can operate any application, exactly the way Manus operates its sandbox. It is available on the Anthropic API and through a growing number of partner products. The developer experience is less polished than Manus’s consumer interface, but the capability is substantially the same.

Claude for Chrome is the more approachable version for a non-technical operator. It is an extension that gives Claude the ability to see and act on the current browser tab — filling forms, extracting data, moving through multi-step flows — with the human staying in the loop and approving steps. It is the sensible default for a UK team that wants an agent experience without the compliance headache.

Crucially, both run under Anthropic’s standard commercial terms, which include a UK-law data-processing addendum, SCCs where relevant, and a clear commitment that paid-tier conversations are not used for model training. For a UK business with client data in the loop, that is the difference between an agent you can defend to a compliance officer and one you cannot.

A practical pilot you can run this week

If, having read all of the above, you still want to understand Manus for yourself — which is the right instinct — here is a non-risky way to do it. Pick one task that is entirely public, entirely non-personal, and that you already know how to do by hand. The classic candidate is an SEO audit of your own public homepage, or a competitive teardown of five companies whose information is all in their public marketing.

Low-risk Manus pilot brief
You are an autonomous research agent. Task: 1. Open the following five competitor websites: [list five public competitors]. 2. For each, extract: pricing (or note "on request"), target customer, positioning statement, three most prominent features, and any reviews from the last 12 months you can find on G2, Trustpilot or Capterra. 3. Produce the output as a comparison table with one row per competitor and one column per attribute. 4. Add a short paragraph at the end flagging where you were uncertain or had to guess. Constraints: - Do not log in to any site. - Do not use any personal data. - Cite the source URL for every claim.

Run that once. Time how long it takes. Count the errors. Compare it against the version you would have produced by hand. You will come away with a clear-eyed view of where the technology is in 2026, without having exposed a single piece of client information.

The honest verdict

Manus is the most impressive proof of concept we have seen for autonomous agents to date. The orchestration is real. The asynchronous execution model is a genuine step forward. The demos, for once, are not entirely vapour. There is a straight line from what Manus does today to the kind of “delegate a task, get back a result” experience that has been promised by AI vendors for a decade.

For UK small businesses in late 2026, however, the compliance posture is not yet defensible for client data. A tool built and operated outside the UK and EU, with no SCCs, no UK data-processing agreement and no UK or EU region, is not one you should be pushing personal data or credentials into. That is not a criticism of the technology. It is a criticism of the legal wrapper around the technology, and it is the kind of thing that will probably improve in the next twelve to eighteen months as Butterfly Effect matures.

Until then, our recommendation is unchanged. Experiment with Manus on public data to understand what agentic AI looks like. Run your production workflows — the ones that touch your clients, your CRM and your inbox — through Claude, whose UK legal footing is clear and whose agent features now cover most of the same ground. When Butterfly Effect ships UK data residency and a signable UK DPA, revisit. Until then, the answer for most of our readers is: interesting, worth watching, not safe for your client data today.