Manus went viral in 2025 in an unusual way. Invite codes were selling for thousands of dollars. What people were chasing was its agent capability: give it one sentence, and it handles the rest. Not just a chat assistant. Something that actually does things.
Two years on, it’s open to most users for free. The conversation around it has also gotten more complicated: company relocation, a Meta acquisition that China then ordered unwound, its return to independent operations in August 2026, and data privacy questions. This article tries to lay out what it can and can’t do, plus the company background, and let you decide if it fits.
What Manus Is
Manus (manus.im) is an autonomous AI agent built by Butterfly Effect Pte. Ltd. The key difference from a regular AI chat tool is execution: give Manus a task and it breaks it into steps, then goes and browses pages, runs code, manages files, and produces a finished output. That might be a research report, a presentation, organized data, or a website prototype. You don’t need to guide it one prompt at a time.
You can watch the whole process live in the right-side panel: which page it’s checking, what terminal commands it’s running, which step it’s on. That transparency is one of the things Manus specifically designed for.
Manus doesn’t train its own base language model. It relies on third-party models. Based on public reporting, it primarily uses the Claude family as its reasoning backbone, which means a Claude outage affects Manus availability too.
Company Background and the Controversy
A few facts worth knowing, taken together.
Origins and relocation. The Manus development team was founded in China (Beijing and Wuhan) in 2022. By mid-2025, the company moved its legal entity to Singapore, shut down its mainland team, and closed its Chinese social accounts.
The Meta acquisition. Meta completed its acquisition of Manus in December 2025, reportedly for around US$2 billion. On April 27, 2026, China’s National Development and Reform Commission (NDRC) issued an order requiring the completed deal to be unwound, citing national security. The NDRC’s position: Singapore registration doesn’t place the company outside Beijing’s jurisdiction. This was the first time China used foreign investment security review to unwind a completed cross-border AI transaction.
Return to independent operations. On August 11, 2026, Manus announced it would separate from Meta and return to operating as an independent company (Bloomberg). On September 1, the official blog announced that it had formally resumed independent operations, led by the founding team. Reports also said the founders had explored raising around $1 billion to buy the company back; the final ownership arrangement has not been disclosed by the company.
Data deletion in August 2026. As part of the separation and to meet requirements in certain jurisdictions, an official notice said that data created by some users on or after December 29, 2025 was deleted on August 23–24, 2026 (Singapore time). Affected users could back up their data before August 23 and restore it from the recovery page starting August 25; the company says restoration has no deadline. If you used Manus during this period, sign in and check that your data is intact.
Data privacy. Manus is legally based in Singapore, but the team’s background and early operations have extensive China connections. The company says that after its return to independence, data is stored in the United States and Singapore (official notice); more detailed data flows have not been fully disclosed. If your use case involves sensitive business data or personal information, that context is worth a minute of thought before you start. No evidence says there’s a problem. The public information also isn’t sufficient to rule one out. That’s the honest position.
These are verifiable facts. People weigh them differently. This article doesn’t draw the conclusion for you.
How to Get Started
- Go to manus.im
- Sign in with email, Google, Apple, or Microsoft
- After account setup you’ll receive starting credits (roughly 1,000-1,500)
- Select Agent Mode (not Chat Mode: Chat Mode is free but doesn’t execute tasks)
- Attach any relevant files as background context
- Write your task prompt: include context, goal, expected output format, and a numbered list of requirements
- Click Run and watch execution in the right panel
- Download the output when done; keep an eye on the credit balance in the top right corner throughout
Write specific prompts. Vague instructions make the agent spend more steps figuring out what you want, and every step burns credits. “Research the Taiwan electric vehicle market and produce a PDF report in Traditional Chinese with market share, major brands, and consumer trends, with source citations on the last page” is not just better quality than “research the EV market.” It’s cheaper and takes fewer retries.
The “My Computer” feature released in March 2026 lets Manus access local files and the terminal. It asks for permission command by command; it doesn’t operate silently in the background.
Pricing and Credits
| Plan | Monthly | Credits | Best for |
|---|---|---|---|
| Free | $0 | 300/day | Testing, occasional use |
| Pro | From $20 | From 4,000/month | Regular use |
| Pro | From $40 (7-day free trial) | From 8,000/month | Medium-high frequency |
| Team | From $20 per seat | Shared team credits | Small teams |
Based on Manus’s official plan information (updated 2026-03-16). Annual billing saves 17%, and Pro has higher-credit options. Plans change frequently; check manus.im pricing for current details.
The important credit warning:
Chat Mode is free. Agent Mode costs credits. Credit consumption varies a lot: simple queries run 10-50, deep research tasks 500-900.
Here’s the trap: independent reviews consistently show actual consumption running well above Manus’s own estimates, in some cases 2-4 times higher. A complex open-ended task can burn through an entire month’s allowance in a single run.
A few mechanics to know:
- Free plan daily credits don’t carry over. What you don’t use today is gone.
- Plan monthly quotas don’t roll over. They reset each billing cycle. Separately purchased credits don’t expire, as an exception.
- If credits run out mid-task, the agent stops where it is. You get an incomplete output, with no graceful recovery.
The practical move: before running a new type of task at full scale, run a smaller version first to see how fast it burns, then decide whether to go bigger.
What It’s Good at and Where It Falls Short
Where Manus works well:
- Open-ended multi-source research. Researching a market, building a competitive analysis, pulling together information from many sources. This is where Manus is strongest.
- Travel and itinerary planning. Give it a destination, number of days, and budget. It can check flights, attractions, and accommodation and assemble something usable.
- Long tasks you can walk away from. Fifteen to ninety minutes of background execution without babysitting.
- Watching AI work. If you want to see an AI step through a process in real time, Manus has the most transparent interface for it.
Real limitations (from independent reviews):
- Paywalls and CAPTCHAs stop it cold. Sites requiring login or verification checks trip it up constantly.
- Stability is inconsistent. Crashes, high-load errors, and mid-task stops are reported by many users, not a handful.
- Coding on a real codebase is weaker than Cursor or Claude Code. Editing existing projects or debugging is not where Manus shines.
- Credits running out mid-task means an unfinished result. There’s no graceful pause-and-resume.
- No team collaboration features on basic plans.
Short version: Manus has real value for its best scenarios (long tasks, open-ended research, data aggregation). It’s not a universal agent. Framing it as “research assistant” rather than “does everything” sets more accurate expectations.
Traditional Chinese and Taiwan Access
Traditional Chinese. The underlying Claude-family models should handle Traditional Chinese input at a reasonable level. Manus hasn’t officially listed Traditional Chinese as a tested or supported language, and the interface is primarily in English. Personal testing is the only reliable gauge. Specifying “Traditional Chinese” output in your prompt usually helps.
Taiwan access. Taiwan works. Manus runs a Traditional Chinese (zh-tw) page and is actively marketing to Taiwanese users. The only confirmed geo-block is mainland China IP addresses. Claims circulating online about Taiwan being restricted don’t trace back to a reliable source; the likely explanation is confusion with restrictions on other China-origin AI products.
The data privacy reality. Manus operates under Singapore law. As covered in the background section, data is stored in the United States and Singapore, but more detailed data flows remain unclear. If your task involves sensitive business information, personal data, or anything that needs to stay confidential, spend a moment evaluating that before you proceed. Treat it as standard due diligence before picking any tool.
How It Compares to Other Agents
| Manus | ChatGPT Agent | Claude (direct) | |
|---|---|---|---|
| Long autonomous tasks | Strong | Average | Weak (not designed as an agent) |
| Short-task efficiency | Burns credits fast | Good value (included in $20/month) | Depends on plan |
| Coding on real projects | Weaker than Cursor/CC | Average | Strong (especially Claude Code) |
| Execution transparency | High | Average | Low (no live execution panel) |
| Stability | Inconsistent | More stable | Stable |
| Underlying model | Uses Claude as backbone | OpenAI GPT | Claude is the product itself |
ChatGPT Agent is better value for most users: it comes with Plus at $20/month, and short tasks don’t drain it. Manus’s differentiation is specifically in long open-ended research tasks. Claude itself is stronger than Manus for document analysis, long-form writing, coding, and API use, but doesn’t have Manus’s autonomous multi-step execution (Claude Code is a different category, not the same type of product).
OpenAI Deep Research is the closest competitor to Manus on research use cases. Both aggregate from multiple sources. The difference: Deep Research output skews more academic, Manus is more dynamic in execution.
Conclusion
The problem Manus solves is real. Many tasks actually need execution power, not just smart answers. A tool that can run fifteen steps on its own has value. Manus does that part.
But a few things are worth evaluating before upgrading.
Credit burn needs a calibration period. Budget 2-4x your expected spend and you’re less likely to be surprised. The Free plan’s 300 daily credits is fine for trying one or two tasks. As an actual work tool, you need to do the math.
Stability is a genuine problem. Paying for a task that stops halfway through is frustrating. Multiple independent reviewers have flagged this. It’s not an edge case.
The company background is a fact, not a verdict. Singapore relocation, the Meta deal being unwound followed by a return to independent operations, and limited data transparency: these are checkable. How much weight they carry depends on what you’re using Manus for. Dropping in personal entertainment tasks is a different conversation from putting confidential company data through it.
Manus is a good fit if you want to hand off long multi-source research, you’re willing to learn the credit system, and your tasks don’t involve highly sensitive data. Less of a fit if your main need is coding, you have low tolerance for tool instability, or you need to process anything sensitive.
Official documentation: help.manus.im
Further Reading
Manus plans and features change often. Check manus.im for current information.
By Penna | 小企鵝 Penchan