Manus is a general-purpose autonomous AI agent product developed by Butterfly Effect, a company founded in China and later headquartered in Singapore. Publicly unveiled in March 2025, it was marketed as a system that carries out multi-step digital work, such as research, data analysis, coding and website building, with limited ongoing human direction. The name comes from the Latin word for "hand," meant to suggest the ability to act rather than only to advise.
Origins and development
Butterfly Effect was founded in 2022 by Xiao Hong, shortly before the public release of chatgpt. Its earlier consumer product, Monica, was a browser extension and all-in-one assistant launched in 2023 that gave users access to several large language models from a single sidebar. According to accounts of the company's history, development of Manus as a standalone agent began in late 2024, when the team concluded that the next phase of consumer AI would involve systems that act on a user's behalf rather than only generating text.
Sources differ on the exact launch date. Most give March 6, 2025 as the public release, while some describe a March 5 demonstration. Access was initially invitation-only, and one industry profile reports that invitation codes were resold at high prices in China. In April 2025 the company raised a $75 million round led by the U.S. venture firm Benchmark, and in June 2025 a co-founder confirmed that headquarters had moved to Singapore.
Architecture and functions
Manus runs each task inside a cloud-hosted virtual environment. Its documentation describes a sandbox with internet access, a persistent file system and the ability to install software and create custom tools, which distinguishes it from chat interfaces that only return text. Tasks continue to run asynchronously, so users can check results later.
Public descriptions characterize the system as multi-agent. A planning component breaks a request into subtasks and delegates them to specialized sub-agents for functions such as knowledge retrieval and code generation, which can run in parallel. Some profiles also describe browser automation that can act within a user's authenticated sessions. Typical outputs include research reports, spreadsheets such as CSV or Excel files, slide decks and deployed web pages.
Manus is model-agnostic. Rather than relying on one proprietary model, it orchestrates third-party models, and public accounts cite Chinese and Western providers including Anthropic's Claude and Alibaba's Qwen. The company has reported partnerships with Alibaba on a Chinese-language version and on open-source models, though the technical details of these arrangements are not fully public.
Benchmark claims and reception
At launch, the company stated that Manus achieved state-of-the-art results on the GAIA benchmark, a test of general AI assistant capabilities that includes questions requiring tool use and multi-step reasoning. Reported scores across the three difficulty levels were about 86.5%, 70.1% and 57.7%. These figures originate with the company and were widely repeated in coverage; the entry does not treat them as independently verified.
Early reviews were mixed. A hands-on test by MIT Technology Review found frequent system crashes and server overload during the first week of public access, while judging the interface intuitive and the approach promising. Commentators compared the launch with the surprise impact of DeepSeek in early 2025, and some scientists described the system as a possible early glimpse of artificial general intelligence, a characterization that remains contested.
Commercial growth and Meta acquisition
The company stated in December 2025 that its annual recurring revenue had passed $100 million about eight months after launch. Meta later said the service served millions of users and businesses. On 29 December 2025, Meta Platforms announced an agreement to acquire Manus. The Wall Street Journal reported a price above $2 billion, and other estimates ranged up to $2.5 billion. Meta stated that it would continue operating and selling the Manus subscription service, integrate its technology into products such as Meta AI, and wind down Manus's remaining business in China, with no continuing Chinese ownership. The company's chief executive, Xiao Hong, was to join Meta.
Regulatory reversal and independence
The deal drew scrutiny from Chinese authorities. According to reports, the Ministry of Commerce opened a review in early 2026, and on 27 April 2026 the National Development and Reform Commission issued a decision prohibiting the investment and requiring the transaction to be rescinded, with the parties restored to their pre-transaction positions. In August 2026, Manus announced that it would resume operations as an independent company and separate from Meta. On 8 October 2026, Butterfly Effect said in a WeChat post that it had raised more than US$500 million in a round led by Boyu Capital and IDG Capital, with existing investors including Tencent, HongShan and ZhenFund participating. This was the company's first round since the Meta acquisition was abandoned. One deal database reports a valuation of about $4 billion for this round; other sources did not confirm the figure.
Sources give somewhat different chronologies for the independence announcement and the financing, so the dates above reflect the most consistently reported points.
Analysis, controversies and limitations
Commentators read the Meta deal as an example of value in AI shifting toward the "execution layer," meaning systems that turn model outputs into completed work, rather than toward model development alone. One analysis noted that Manus did not present itself as a novel model maker; its distinctive contribution was how it combined existing models into a usable product.
The episode also became a case study in the US–China technology rivalry, with the acquisition seen as an unusual instance of a U.S. company buying a Chinese-founded firm and then being blocked by Beijing. More broadly, agents that operate inside users' accounts, browse the web and run code raise concerns about data privacy, security and regulatory compliance, which are cited as obstacles to wider adoption. Early performance problems, such as outages during peak demand, also illustrate the reliability challenges common to agentic systems.
Claims that Manus was the first general-purpose agent remain disputed, since earlier systems and other agent frameworks exist. The product is best understood as a widely publicized commercial implementation of the multi-agent approach within the broader movement toward generative AI systems that act autonomously.
References
- Manus (AI agent)en.wikipedia.org
- Welcome - Manus Documentationmanus.im
- Introducing Manus: The general AI agent — WorkOSworkos.com
- Manus AIaiwiki.ai
- Manus - Learn AIai.miraheze.org
- Butterfly Effect Technology: Manus AI Founders, Funding, Meta Deal Timelinejademond.com
- Manus（The world's first general-purpose AI assistant.）_Baiduwikibaike.baidu.com
- Everyone in AI is talking about Manus. We put it to the test.technologyreview.com
- China's Manus AI 'agent' could be our 1st glimpse at artificial general intelligencelivescience.com
- Meta acquires intelligent agent firm Manus, capping year of aggressive AI movescnbc.com
- Meta acquires Chinese-founded Manus for over $2bn in AI pushsiliconrepublic.com
- Meta to acquire AI agent company Manusfinance.yahoo.com
- Manus AI Acquisition, Valuation, Funding & Investorsmultiples.vc
- Manus Raises $500M+ as an Independent Company After Meta Deal Collapses - WOWTALEen.wowtale.net
- Manus came home, and Singapore couldn't stop ite27.co
- Manus AI Agent: What It Is, How It Works, & Its Impact [2025]leanware.co
- Butterfly Effect (Manus) in 2026: Owner, Funding, Founders & the Meta Dealtheairankings.com
- [2505.02024] From Mind to Machine: The Rise of Manus AI as a Fully Autonomous Digital Agentarxiv.org