Manus AI
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v8 · 6,282 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v8 · 6,282 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: artificial intelligence terms
Manus AI is an autonomous artificial intelligence agent publicly unveiled on March 6, 2025, by the Chinese startup Butterfly Effect (also known as Monica.im).[1][40] Named after the Latin word for "hand," Manus was built to bridge the gap between human intent and real-world action, independently executing complex tasks with minimal supervision.[8] It became one of the most discussed AI agent products of 2025, generating viral attention on social media within hours of its launch.[6] Manus describes itself as "the first General AI Agent," a category of agentic AI that takes actions on a user's behalf rather than only generating text.[20] By December 2025 the company said the product had processed more than 147 trillion tokens and created more than 80 million virtual computers since launch.[20]
The product runs primarily on Claude from Anthropic and fine-tuned Qwen models from Alibaba rather than a proprietary foundation model.[33] It operates inside a sandboxed Linux environment with a browser, terminal, and file system.[37] After raising $75 million from US venture firm Benchmark Capital in April 2025 at a valuation of around $500 million, Butterfly Effect relocated its operations from China to Singapore.[9][10] By December 2025 the company announced it had reached roughly $100 million in annualized recurring revenue and a run-rate exceeding $125 million when usage-based revenue was included.[20] In December 2025, Meta announced an agreement to acquire the company at a valuation in the range of $2 billion to $3 billion.[23][24] China's National Development and Reform Commission blocked the deal on April 27, 2026.[25][26]
Manus is marketed as a "general" AI agent that does more than answer questions: it browses the web, writes and runs code, manages files, fills out forms, and produces deliverables such as reports, websites, slide decks, and spreadsheets.[1] Users submit a natural-language task and the agent runs asynchronously in the cloud, notifying the user when it is finished or when it needs a decision. This distinguishes it from chat assistants such as standard ChatGPT sessions and from interactive browser agents like OpenAI Operator, which require the user to keep the page open while the agent works.
The core product combines three components: a reasoning model (the Claude family), a fine-tuned secondary model (the Qwen family) used for cheaper or specialized tasks, and an orchestration layer that exposes a Linux sandbox and a Chromium browser to the agent.[33] Inside that loop the agent reads, plans, calls tools, observes the result, and iterates until the task is complete or stuck.
Yichao Ji, known by the English name Peak Ji, was born in 1992 and spent his early childhood in Boulder, Colorado, where his father held a visiting scholar post.[13] The family returned to Beijing when Ji was in second grade. He grew up bilingual, taught himself operating system development in high school while running the school computer club, and traces his programming interest to an elementary school robotics team.[13]
In 2010, as a high school sophomore, Ji released Mammoth Browser, a customizable iPhone browser that became the most-downloaded third-party browser by an individual developer in China. It won the Macworld Asia Grand Prize in 2011 and earned Ji a spot on Forbes China's 30 Under 30 list at age 20.[13] He followed it with other iOS apps, including a Monopoly-branded budgeting tool that attracted a trademark notice from Hasbro, an early lesson in product risk that he later credited with sharpening his judgment.
In 2012, Ji founded Peak Labs with backing from ZhenFund and HSG (then known as Sequoia China) and led development of Magi, a search engine that extracted information from across the web to answer queries in a way conceptually similar to today's AI-powered search, but built on a custom language model.[13] Magi gained millions of users early but failed to sustain consumer adoption. Ji pivoted to B2B applications before selling the product in 2022.[13]
MIT Technology Review named Ji one of its 2025 Innovators Under 35 for his work on Manus.[14]
The direct predecessor to Butterfly Effect was Monica, also stylized as Monica.im. Xiao Hong (also written Hong Xiao), a 2015 graduate of Huazhong University of Science and Technology in Wuhan, founded Butterfly Effect in 2022, two months before OpenAI publicly launched ChatGPT. The company kept offices in Beijing and Wuhan and initially targeted consumers in North America, Japan, and South Korea.
In March 2023, Butterfly Effect launched Monica.im, an all-in-one AI personal assistant and Chrome browser extension that aggregated multiple large language models behind a single sidebar interface for translation, summarization, and writing assistance. The extension reached tens of millions of installs and gave the team experience shipping consumer AI to a global audience.
The founding team included three principals: Xiao Hong as CEO, Ji Yichao as co-founder and chief scientist leading technical work, and Zhang Tao as co-founder and product director. Zhang Tao directed the launch demo video that went viral in March 2025 and became a key part of the public narrative around Manus.
In 2024, the Monica team began experimenting with autonomous agents, betting that the next consumer wave after chatbots would be agents that took actions on a user's behalf rather than just generating text. They prototyped an internal agent that used the Anthropic and Alibaba APIs to drive a sandboxed browser and terminal, then refined it through several internal iterations before deciding it was ready to show publicly. The agent product was branded Manus while the parent company took the name Butterfly Effect.
The early prototyping relied on open-source infrastructure including the Browser Use library for browser automation and Playwright for headless Chromium control, alongside Microsoft Azure cloud services and Anthropic's Claude API.
Manus was unveiled on March 6, 2025, through a five-minute demo video featuring Peak Ji walking through several real tasks: building a stock-research dashboard, generating a Japan travel handbook, screening resumes against a job description, and producing an interactive middle-school physics lesson.[40] Access was invite-only from day one.[5]
The launch went viral almost immediately. The demo video amassed over 200,000 views in 24 hours, with some reports citing over one million views within twenty hours.[5] The official Discord server passed 138,000 members within a week. Invitation codes were resold for the equivalent of thousands of dollars on Chinese resale platforms such as Xianyu and on international sites such as eBay.[3] Within a week the waiting list reportedly reached around 2 million people.
Coverage from outlets including the South China Morning Post, MIT Technology Review, Business Insider, Forbes, TechCrunch, and Quartz framed Manus as the next big moment in Chinese AI after DeepSeek, which had captured global attention in January 2025.[4][7] Rowan Cheung of The Rundown AI newsletter called it "China's second DeepSeek moment."[2]
Butterfly Effect was unprepared for the volume of demand. Product partner Zhang Tao publicly attributed the access bottleneck to capacity rather than a deliberate scarcity strategy: "The current invite-only mechanism is due to genuinely limited server capacity at this stage." Critics, including Pleias co-founder Pierre-Carl Langlais, argued that the rollout amounted to "hunger marketing" that artificially inflated hype.[3]
In May 2025, Butterfly Effect opened Manus to the public through a paid subscription model, ending the invite-only phase.[21]
Unlike DeepSeek, which trained its own foundation model, Butterfly Effect did not build a frontier large language model. Manus instead orchestrates external models. Peak Ji has confirmed that Claude 3.5 Sonnet from Anthropic is the primary reasoning engine, with fine-tuned versions of Alibaba's Qwen models used for cheaper or auxiliary calls.[33][37] Subsequent versions began experimenting with Claude 3.7 Sonnet for harder reasoning tasks.
This approach is sometimes described as a "wrapper" architecture, but the engineering is non-trivial because the models are integrated into a runtime that gives them a sandboxed Linux environment, a Chromium browser, and a memory and file layer that persists across long-running tasks.[34] Internal testing has also reportedly explored multi-model routing, directing specific subtasks to the model best suited for them.
Manus follows the CodeAct pattern, in which an agent emits Python code rather than structured JSON tool calls. Each step of the loop the model writes a short Python program that uses the available tools, the runtime executes that program inside the sandbox, and the resulting stdout, screenshots, and side effects are folded back into the model's context for the next step. The advantages of CodeAct are that the agent can compose tools fluidly, maintain state across steps in ordinary variables, and handle conditional logic without round-tripping every branch through the language model.
The system prompt that governs the agent is organized into modular tagged sections including <tool_use_rules>, <browser_rules>, <coding_rules>, and <writing_rules>. Key directives include one action per iteration before observing results, a prohibition on direct natural-language responses in place of tool invocations, prioritization of authoritative APIs over general web scraping, mandatory source citation for factual claims, and self-directed error diagnosis and retry with alternative approaches.[34]
The orchestration layer wraps this loop with several specialized sub-modules:
| Module | Role |
|---|---|
| Planner | Breaks the user's request into a structured plan and a numbered todo list |
| Executor | Runs the CodeAct loop against the sandbox and browser |
| Knowledge retriever | Fetches reference material from the web or from prior steps |
| Memory | Persists notes, files, and intermediate artifacts between steps |
| Observer | Captures screenshots, command output, and file changes for the model |
| Event stream | Chronological log of all interactions, truncated or summarized for context length management |
The official Manus documentation lists 29 integrated tools and open-source utilities that the agent can call inside the sandbox, including a Chromium browser controlled through the Playwright protocol, a shell, file system commands, image and PDF processing, and deployment helpers that can publish a finished website or report to a public URL.[34]
Each Manus session runs inside an isolated Linux container in the cloud. The agent gets its own file system, its own browser profile, and a fresh execution environment. Sessions are stateless across users, keeping one user's data from leaking into another's session. Users can watch the agent work in real time through a replay view that shows each browser action and shell command as it happens.
State management uses files as the primary persistence layer. The agent writes a todo.md file to track which steps it has completed, saves intermediate data to domain-specific files in the workspace, and uses a retrieve-augmented approach to pull external documents and data into context rather than relying on parametric memory.
For complex projects, Manus can deploy specialized sub-agents in isolated sandbox environments, with one handling web research, another coding, and a third managing data analysis. A high-level orchestrator divides tasks and integrates results, enabling parallel work on different aspects of a single user request. This architecture was the foundation for the Wide Research feature introduced in July 2025.
On July 31, 2025, Butterfly Effect launched Wide Research, a feature that deploys over 100 parallel agent instances simultaneously for large-scale data-gathering and analysis tasks.[15] The launch made Manus one of the first consumer agent products to operationalize massively parallel multi-agent execution at that scale.
Wide Research differs from sequential or single-agent research in a fundamental way: rather than processing items one after another in a growing context window (which causes quality degradation as the list grows), each sub-agent operates independently with its own virtual machine, browser, tools, and fresh context.[17] The official description calls it "fresh context for every item" rather than the accumulating context that causes traditional AI systems to produce "generic filler" by item 50.[17]
The architecture passes through four stages:
In a public demo by Peak Ji at launch, Wide Research spun up exactly 100 agents, each assigned a different sneaker model, and produced a sortable comparison matrix covering design, price, availability, and category in under five minutes.[15] Other demonstrated use cases include profiling 250 AI researchers with publication records, generating company competitive intelligence reports across large vendor lists, and simultaneously creating batches of images with consistent visual themes.[17]
Wide Research initially launched for Pro tier subscribers and later became available to Plus and Basic plan users.[15] Bloomberg covered the launch under the headline "OpenAI-Challenger Manus Preps Big Upgrade to Main Agent Platform."[16]
On November 22, 2025, Butterfly Effect released Manus Browser Operator to all users.[18] Unlike the default Manus setup, which drives a cloud-hosted Chromium browser the user never sees, Browser Operator is a browser extension that connects the Manus agent directly to the user's own local browser session.[18]
The practical significance is access to authenticated services. A user logged into Crunchbase, PitchBook, SimilarWeb, Semrush, or a proprietary CRM can point Manus Browser Operator at those platforms without sharing credentials or dealing with CAPTCHA interruptions, since the agent operates from the user's existing session at their trusted local IP address.[18] The extension provides a full audit trail of all actions and can be stopped instantly by closing the tab.[18]
Browser support at launch was Chrome and Edge, with additional browsers listed as forthcoming.[18] The product works alongside, rather than replacing, the cloud browser automation that Manus uses for public websites.
| Date | Milestone |
|---|---|
| March 6, 2025 | Invite-only public launch; demo video goes viral |
| April 2025 | $75M Series B led by Benchmark Capital |
| May 2025 | Public access opened; first paid subscription tiers introduced |
| June 2025 | Singapore headquarters officially listed; co-founders relocate |
| July 31, 2025 | Wide Research launched for Pro users (100+ parallel agents) |
| September 2025 | Peak Ji named MIT Technology Review Innovator Under 35 |
| October 22, 2025 | Manus 1.5 launched (faster speed, higher quality, unlimited context, upgraded builder) |
| November 22, 2025 | Manus Browser Operator launched for all users |
| December 15, 2025 | Manus 1.6 launched (1.6 Max mode, Mobile Development, Design View) |
| December 17, 2025 | Company announces $100M ARR and $125M run-rate |
| December 29, 2025 | Meta announces acquisition agreement |
| April 27, 2026 | China's NDRC blocks the Meta acquisition |
At launch Butterfly Effect claimed state-of-the-art performance on GAIA, a third-party benchmark for general AI assistants developed by researchers at Meta and Hugging Face.[39] GAIA scores agents across three difficulty levels covering real-world reasoning, web research, and tool use. The numbers Manus reported at launch were as follows.
| GAIA difficulty | Manus | OpenAI Deep Research | Previous SOTA |
|---|---|---|---|
| Level 1 (basic) | 86.5% | 74.3% | 67.9% |
| Level 2 (intermediate) | 70.1% | 69.1% | 67.4% |
| Level 3 (complex) | 57.7% | 47.6% | 42.3% |
Butterfly Effect said the runs used the same configuration as the production product and were therefore reproducible by anyone with access.[39] The figures were not independently audited at the time of launch, and several reviewers questioned whether the production system actually matched the benchmark configuration. TechCrunch reporter Kyle Wiggers documented loops, crashes, and refusals on basic requests that the benchmark numbers would not predict.[2]
Despite the skepticism, the GAIA results were influential. They made Manus the first widely publicized agent product to claim a clear lead over OpenAI's Deep Research on a public benchmark, and they contributed to the framing of Manus as a possible second DeepSeek moment for Chinese AI.[31]
Butterfly Effect markets Manus as a general agent. Demonstrated and supported workloads include the following.
| Use case | Example |
|---|---|
| Travel planning | Generate a multi-day Japan itinerary with bookings, maps, and a printable handbook |
| Financial analysis | Pull stock fundamentals, build an interactive dashboard, and write a written summary |
| Resume screening | Read a folder of PDFs, score candidates against a job description, and rank them |
| Educational content | Build interactive lessons such as a middle-school physics module on the momentum theorem |
| Document comparison | Diff long contracts or insurance policies and produce a side-by-side table |
| B2B sourcing | Search for suppliers, gather contact details, and assemble a vendor shortlist |
| E-commerce analytics | Pull store data, compute trends, and generate optimization recommendations |
| Web development | Build and deploy a small static or React website from a brief |
| Data analysis | Clean a CSV, run pandas analyses, and ship a chart-laden report |
| Mobile development | Build mobile applications from natural-language descriptions (added in Manus 1.6) |
| Parallel research | Profile hundreds of companies, researchers, or products simultaneously via Wide Research |
A distinguishing feature compared to interactive agents like Operator is the asynchronous workflow. The user submits a task and is free to log off; Manus continues to work in the cloud and notifies the user when the result is ready or when it needs more input. The agent can also publish artifacts to a public URL so the user can share results without first downloading them.
Manus initially launched as invite-only with no pricing. In May 2025 Butterfly Effect introduced paid tiers built around a credit system, where each task consumes a number of credits depending on its length and the tools it uses.[21] The pricing structure evolved through several revisions over 2025.
By 2026, the current tiers are structured as follows.[22]
| Plan | Monthly price | Monthly credits | Key access |
|---|---|---|---|
| Free | $0 | 1,000 starter credits; 300 refreshed daily | Manus 1.6 Lite in Agent Mode, Chat Mode |
| Pro (Standard) | $20 | 4,000 | Manus 1.6 Lite, 1.6, and 1.6 Max in Agent Mode |
| Pro (Extended) | $40 | 8,000 | Full access including Max mode; 7-day free trial |
| Pro (High Volume) | $200 | 40,000 | High-volume individual use |
| Team | $20/seat | Varies by seat count | SSO, analytics, access controls, shared templates |
The credit system attracted criticism for being opaque. Reviewers noted that two superficially similar tasks could consume very different amounts of credit depending on how many browser steps or how much code execution the agent decided to run, making it hard for users to predict monthly costs. Butterfly Effect later added in-product cost estimates and clearer per-task credit breakdowns in response to the feedback.
Credits do not roll over between months for base plan allocations, though separately purchased add-on credit packs carry over while a paid subscription remains active.
Manus moved through several funding rounds during 2025.
| Date | Round | Amount | Lead investor | Reported valuation |
|---|---|---|---|---|
| Pre-launch (2024) | Seed | ~$10M+ | Tencent, HSG (Sequoia China successor), ZhenFund | Undisclosed |
| April 2025 | Series B | $75 million | Benchmark Capital | ~$500 million |
| December 2025 | Acquisition agreement | Reported $2-3 billion | Meta Platforms | $2-3 billion |
The Benchmark-led $75 million Series B in April 2025 (reported by Bloomberg and others; it followed an earlier round of north of $10 million) was Benchmark's first Chinese AI investment and was widely covered as a sign that top-tier US venture firms were willing to bet on Chinese-founded AI startups despite geopolitical headwinds.[9][10] Benchmark general partner Chetan Puttagunta joined the board.
In May 2025, the US Treasury Department issued an inquiry under the Outbound Investment Security Program (OISP), a framework established by President Biden's 2023 executive order that took effect in January 2025.[11][12] The OISP requires US entities to notify Treasury of investments in key sectors, including AI, that could advance sensitive technologies in countries of concern.[11]
Benchmark's legal counsel argued the investment did not require notification on three grounds: Manus was a "wrapper" that used existing AI models rather than developing its own; the parent company was incorporated in the Cayman Islands; and employees worked across the US, Singapore, Japan, and China rather than in China exclusively.[12] Data was stored on Western cloud infrastructure including Microsoft Azure. As of May 2026, the review had not produced a public enforcement action.
The round attracted criticism within Silicon Valley. Some argued that investing in a Chinese-founded AI company with significant China-based operations was incompatible with US national security interests regardless of corporate structure. Others pointed out that the same reasoning would apply to dozens of other enterprise software companies with distributed global workforces.
Following the Benchmark round, Butterfly Effect relocated its headquarters from Wuhan and Beijing to Singapore.[30] By June 2025, the company had updated its website to list Singapore as its global headquarters and incorporated a new Singapore-based entity to operate the AI agent product in markets outside China. The three co-founders relocated in mid-2025.
Of approximately 120 China-based employees, around 40 core technical staff relocated to Singapore. The remainder were laid off as part of what the company described as "optimizing its domestic business team."[30]
The relocation served several purposes. Singapore has no domestic AI regulation that would restrict development of agentic AI products. It gave Butterfly Effect cleaner access to US cloud infrastructure and reduced the political friction involved in marketing to US enterprise customers. It also helped position the company for the Benchmark investment, since the firm's lawyers could argue the company was not technically Chinese-based.
Critics, primarily Chinese business press and nationalist commentators, described the move as an attempt to use Chinese engineering talent and capital to build value offshore, then monetize that value through a Western acquisition structure that bypassed Chinese regulatory oversight. The term "Singapore washing" gained currency in coverage of the episode.
Manus sat at the intersection of two concurrent sets of regulatory anxieties in 2025: US concern about Chinese-linked AI companies receiving American capital and using American cloud infrastructure, and Chinese concern about advanced AI technology and talent leaving the country.
On the US side, the OISP review of the Benchmark investment was the most concrete formal action.[11] Legal experts debated whether Manus's fine-tuning work and orchestration engineering constituted "advancement" of AI that would trigger notification requirements, or whether the wrapper characterization put it outside the rule's scope.[12] No enforcement action was announced through April 2026.
On the Chinese side, the NDRC's decision in April 2026 to block the Meta acquisition was the most significant regulatory event in the company's history. The NDRC cited concerns over the cross-border transfer of technology and personnel.[25] Co-founders Xiao Hong and Peak Ji were reportedly barred from leaving China during the review period.
Law firm analyses published after the NDRC decision noted that China's foreign investment security review mechanism, which the NDRC used in this case, has historically focused on defense and energy sectors but is increasingly applied to AI, data, and digital platforms.[29] Analysts at Morgan Lewis described the Manus case as "China's first AI security review block" and noted that the mechanism is more flexible and less transparent than the US CFIUS process, with fewer published rules and precedents.[29]
Foreign Policy and the Asia Times published analyses arguing that Beijing's message was that Singapore incorporation does not shield Chinese-founded AI companies from Chinese sovereignty claims over their technology and talent.[27] The NDRC decision was described as establishing that the "look-through" approach, focusing on origin of technology and source of talent rather than corporate registration, would apply to AI deals going forward.[28]
On December 29, 2025, Meta Platforms announced an agreement to acquire Manus.[23][24] The Wall Street Journal reported the deal at over $2 billion; other reports placed the value as high as $3 billion.[24] Meta framed the acquisition as a core piece of its agentic AI strategy and a way to acquire a working consumer agent product without building one from scratch.[23] Manus had reportedly been in discussions to raise fresh funding at a $2 billion valuation when Meta's offer arrived; the team chose the acquisition path instead.[24]
The Chinese government opened a national security review of the transaction within days. On April 27, 2026, China's National Development and Reform Commission issued a formal block, citing concerns over the cross-border transfer of technology and personnel.[25][26] The official notice stated: "The office in charge of foreign investment security review has decided to block the foreign acquisition of the Manus project and require the parties to unwind the deal."[25]
Legal scholars noted that unwinding an AI company acquisition is structurally different from unwinding a manufacturing acquisition: code, model weights, and engineering knowledge transfer the moment people start working together, and legal reversal cannot undo that transfer. As one analysis from Geopolitechs put it, "you can legally unwind the transaction, but you can't really rewind history."[28]
At the time this article was last updated, Meta and Butterfly Effect had not publicly announced a resolution. Analysts had speculated about restructured arrangements that would leave China-related assets behind.
Manus launched into a crowded field of agent products from US companies. The following table compares Manus to the major agent offerings available in 2025.
| Feature | Manus | OpenAI Operator / ChatGPT agent | Anthropic Computer Use / Claude for Chrome | Google Project Mariner |
|---|---|---|---|---|
| Initial launch | March 6, 2025 | January 23, 2025 (Operator) | October 22, 2024 | December 11, 2024 |
| Underlying model | Claude (Anthropic) plus fine-tuned Qwen (Alibaba) | OpenAI CUA (later GPT-4o based) | Claude 3.5 / 3.7 Sonnet | Gemini 2.0 Flash |
| Foundation model owned by vendor? | No | Yes | Yes | Yes |
| Interaction style | Asynchronous cloud sandbox with browser, terminal, files | Browser session (Operator) or virtual computer (ChatGPT agent) | Sandboxed desktop environment (API) or Chrome extension | Chrome extension on the user's active tab |
| Async background runs | Yes | Limited; ChatGPT agent supports longer runs | Yes (in sandboxed VM) | No |
| Parallel agent execution | Yes (Wide Research: 100+ parallel agents) | No | No | No |
| Public benchmark claims | GAIA L1 86.5%, L2 70.1%, L3 57.7% | OSWorld 38.1%, WebVoyager 87.0% | OSWorld 14.9% (screenshot-only at launch) | WebVoyager 83.5% |
| Pricing entry point | Free tier; $20/month Pro | $20/month ChatGPT Plus (limited agent quota) | API pricing or $20/month Claude Pro | Google AI Ultra subscription |
| Headquarters | Singapore (relocated from China) | United States | United States | United States |
Several patterns stand out. Manus was the only major 2025 agent product that did not own its foundation model, and the only one led by a Chinese-founded team. It was also the most aggressive about asynchronous operation, treating the agent as a cloud worker rather than a copilot riding alongside the user's browser. The Wide Research parallel execution capability had no direct counterpart in the competing products at the time of its launch.
On task style, OpenAI Operator is generally stronger at interactive, real-time browser sessions such as booking tickets or filling out forms, while Manus performs better on longer-running research and multi-step workflows that benefit from running unattended in the cloud.
The initial reaction inside the AI community was a mix of excitement and disbelief. Hugging Face head of product Victor Mustar called Manus "the most impressive AI tool I've ever tried" and said its agentic capabilities "redefine what's possible." Tiny co-founder Andrew Wilkinson said using it to screen job applicants felt like "time travelling six months into the future." AI policy researcher Dean Ball argued that calling Manus a "DeepSeek moment" undersold it because Manus was "actually advancing the frontier" of computer-using AI.
The demo video and the small set of launch use cases drove the bulk of the social-media buzz. Many of the public reactions were second-hand, since invitation codes were scarce, and some early enthusiasts admitted they had not personally used the product when they posted about it.
Within a week of launch the reception cooled as more reviewers got hands-on access. MIT Technology Review's Caiwei Chen wrote a review titled "Everyone in AI is talking about Manus. We put it to the test" that reported a mix of impressive runs and noticeable failures.[3] TechCrunch's Kyle Wiggers documented loops, crashes, and refusals on basic requests.[2] AI researcher Alexander Doria and several others reproduced similar issues. Pierre-Carl Langlais of Pleias accused Butterfly Effect of "deceptive communication" and argued that the GAIA numbers were not consistent with what users actually experienced.[2]
A second wave of criticism focused on novelty. Because Manus did not train its own foundation model and largely orchestrates Claude with a Linux sandbox and a browser, several commentators argued that it was an engineering achievement rather than a research breakthrough. Defenders responded that integration was the hard part: anyone could call the Anthropic API, but very few teams had wrapped that API in a robust agent loop that produced shareable artifacts reliably at scale.
A third strand of criticism concerned data. Researcher Luiza Jarovsky and others raised questions about where Manus stored task data, what was logged, and whether Chinese authorities could access it given Butterfly Effect's then-Chinese ownership. The relocation to Singapore later in 2025 was partly framed as a response to that concern, though critics pointed out that the engineering team remained largely based in China even after the corporate relocation.
Despite the mixed early reviews, the product found substantial paying users. The company reported 20%-plus month-over-month growth following the Manus 1.5 release in October 2025.[32] By December 2025, it announced $100 million in ARR and a run-rate exceeding $125 million when usage-based revenue was included. The company described this as "the fastest startup to go from $0 to $100M in the world."[20] The team at that point numbered 105 people across Singapore, Tokyo, and San Francisco, with a Paris office opening soon.[20] The company said Manus had processed more than 147 trillion tokens and created more than 80 million virtual computers since launch, and that growth had run above 20% month over month since the Manus 1.5 release.[20][32]
Those figures were widely cited as evidence that the market for autonomous agent products was real and growing quickly, even if individual product experiences remained inconsistent.
Butterfly Effect indicated in 2025 that it would open source parts of the Manus stack. Several community-led projects, most notably OpenManus on GitHub, set out to replicate the Manus loop using open-source components, primarily by combining a Claude or Qwen client with a Playwright-driven browser and a CodeAct-style execution loop.[35] OpenManus and similar projects reached tens of thousands of GitHub stars within weeks of the Manus launch, which itself became evidence that the underlying recipe was approachable to any team with API access and engineering bandwidth.[35]
OpenManus was built on the MetaGPT multi-agent collaboration framework. One team member noted they integrated the browser toolchain into the existing code and had the core system running within an hour.[35] The Browser Use library, which Manus itself uses for browser automation, is available as a free MIT-licensed package that developers can self-host or use through a cloud service.
Butterfly Effect's own open source releases were narrower than the community projects and focused on tooling rather than the full agent stack.
Real-world Manus usage has surfaced a recurring set of limitations.
Loops and crashes occur when the CodeAct loop spirals after a website behaves unexpectedly, when a Python script throws an error the model misinterprets, or when the model gets confused about whether a step succeeded. Reviewers reported sessions that burned through hours of agent time without producing a useful result.[2][37]
Cost predictability remains difficult. The credit system makes it hard for users to know what a given task will cost. Two requests that look similar can consume very different amounts of credit depending on how many browser steps or how much code execution the agent decides to run.
Speed is slower than users sometimes expect. Asynchronous cloud execution is convenient, but the agent is slower than a human at most short tasks. A trip-planning request might take ten or twenty minutes; a financial dashboard might take thirty.
Foundation model dependency means that any change to Anthropic's or Alibaba's pricing, rate limits, safety policies, or availability can affect the product directly. This is a structural difference from Operator, Computer Use, or Mariner, all of which run on models their parent companies own.
Tool coverage limits Manus to tasks that can be accomplished through a web browser and command-line tools. It is weaker at tasks that require specialized desktop applications, native mobile apps, or interfaces that are heavily defended against automation.
Privacy concerns persist even after the move to Singapore. Some enterprise users remain cautious about routing sensitive data through a service whose engineering team and source data have Chinese roots, particularly given the broader regulatory turbulence around the Meta acquisition.