Manus (AI agent)

Manus
Screenshot
Developer	Butterfly Effect Technology (Butterfly Effect Pte. Ltd.)
Initial release	March 6, 2025 (2025-03-06)
Stable release	Manus 1.5 / October 16, 2025 (2025-10-16)
Operating system	Web-based, iOS, Android
Platform	Cloud computing
Available in	Multiple languages
Type	Autonomous agent, Artificial intelligence
License	Proprietary
Website	manus.im

Manus (meaning "hand" in Latin) is an autonomous artificial intelligence agent developed by Butterfly Effect Technology, a Singapore-based technology startup with operations in Beijing and Wuhan, China. The agent is designed to independently execute complex real-world tasks without continuous human supervision. Launched on March 6, 2025, Manus has been described as one of the first fully autonomous AI agents capable of independent reasoning, dynamic planning, and autonomous decision-making.

History

Company background

Butterfly Effect Technology was founded by entrepreneur Xiao Hong (Chinese: 肖宏), who previously established Nightingale Technology in 2015. Nightingale developed productivity tools including "Yiban Assistant" (Chinese: 一班助手) and "Weiban Assistant" (Chinese: 微班助手), AI-driven platforms serving over 2 million business users. These products attracted investment from Tencent and ZhenFund.

In 2022, recognizing the potential of large language models, Xiao Hong founded Butterfly Effect and released Monica, an AI assistant browser extension integrating models including ChatGPT and Claude. By 2024, Monica accumulated over 10 million users while maintaining profitability, serving as both a technological foundation and user acquisition platform for Manus.

Development and launch

The Manus project was officially announced on March 5, 2025, on X (social network), with an official launch on March 6, 2025. The founding team includes:

• Xiao Hong (Chinese: 肖宏) – CEO and founder
• Ji Yichao (Chinese: 纪义超), known as "Peak" – Chief Scientist and co-founder
• Zhang Tao (Chinese: 张涛) – Product Partner

Ji Yichao, a former high school dropout, gained recognition for developing Mammoth Browser at age 17.

In early 2024, ByteDance attempted to acquire Monica for US$30 million, but Xiao Hong declined to maintain independence. Monica closed a funding round in late 2024 at an estimated valuation of nearly $100 million.

Following Manus's launch in March 2025, Butterfly Effect raised $75 million in a funding round led by Benchmark at a valuation of approximately $500 million in April 2025. This investment subsequently came under review by the U.S. Treasury Department over compliance with 2023 restrictions on investing in Chinese AI companies.

The product launch featured a four-minute demonstration video showcasing Manus's capabilities. Over two million users joined the waitlist within the first week, causing server load issues and prompting implementation of an invite-only closed beta system. The official Discord community grew to over 138,000 members within days of launch.

Technology

Architecture

Manus employs a multi-agent software architecture with specialized sub-agents working collaboratively:

Planner Agent

Analyzes user requests and creates step-by-step execution plans

Execution Agent

Carries out instructions by interacting with web browsers, databases, and code environments

Knowledge Agent

Handles information retrieval and maintains contextual understanding

Verification Agent

Reviews completed work for quality assurance

This architecture enables parallel subtask processing, improving efficiency compared to monolithic systems. The system operates within a sandboxed cloud-based virtual environment, allowing continuous operation even when users disconnect. Users can close their devices while Manus continues working in the background, sending notifications upon task completion.

Foundation models

According to Chief Scientist Ji Yichao, Manus primarily uses Anthropic's Claude 3.5 Sonnet (later upgraded to Claude 3.7 Sonnet) and fine-tuned versions of Alibaba's Qwen models. The system employs a multi-model approach, dynamically selecting models based on subtask requirements.

CEO Xiao Hong has defended this integration strategy, arguing that value lies in solving user problems through effective integration rather than developing proprietary models, similar to consumer electronics manufacturers using various suppliers' components.

Key features

Asynchronous operation: Manus executes tasks in the cloud while users remain offline, sending notifications upon completion.

Code-Act methodology: Uses executable Python code as its primary action mechanism for complex autonomous operations.

Transparent interface: The "Manus's Computer" feature provides visibility into execution processes, allowing users to review task completion steps through session replay.

Memory and learning: Maintains memory of past interactions and preferences, adapting performance for personalized responses.

Multi-modal capabilities: Processes text, images, tables, and code while generating outputs including reports, visualizations, websites, and spreadsheets.

Performance

GAIA benchmark

Manus achieved notable performance on the GAIA (General AI Assistant) benchmark, developed by Meta AI, Hugging Face, and the AutoGPT team to evaluate AI agents on real-world problem-solving.

Company-disclosed results:

Level	Manus Score	OpenAI Deep Research	Human Average	GPT-4 with Plugins
Level 1 (Basic)	86.5%	74.3%	92%	~15%
Level 2 (Intermediate)	70.1%	69.1%	92%	~15%
Level 3 (Complex)	57.7%	47.6%	92%	~15%

The company claimed state-of-the-art performance across all three tiers, reportedly exceeding OpenAI's Deep Research system. The GAIA benchmark tests fundamental AI abilities including reasoning, multi-modality handling, web browsing, and tool-use proficiency. For context, GPT-4 with plugins achieves approximately 15% on GAIA tests, while human respondents average 92%.

Features and capabilities

Core functionality

Manus distinguishes itself from traditional chatbots and AI assistants by delivering complete task results rather than guidance. Key capabilities include:

Research and analysis: Conducts comprehensive multi-source research on complex topics, synthesizing information into structured reports with citations.

Data processing: Analyzes datasets, creates visualizations, generates statistical summaries, and builds interactive dashboards.

Web automation: Navigates websites, extracts information, fills forms, and performs multi-step online procedures.

Content creation: Produces various content formats including articles, presentations, marketing materials, and technical documentation.

Code development: Writes, debugs, and deploys code across multiple programming languages with integrated testing capabilities.

File management: Processes various file formats including PDF, Excel, CSV, images, and documents, performing extraction, conversion, and analysis.

Use cases

Demonstrated applications include:

• Market research and competitive analysis
• Financial data analysis and stock market screening
• Academic research synthesis
• Travel itinerary planning
• Website and application prototyping
• Resume screening and candidate evaluation
• Business intelligence dashboard creation
• Document automation and processing

Pricing and availability

Access model

Manus initially launched with a closed beta requiring invitation codes distributed through Monica's existing user base and social media channels. Invitation codes were reported to have been resold on black markets for up to ¥50,000 CNY (~US$7,000) during the initial launch period, demonstrating significant demand. The company implemented staged rollout to manage server capacity and gather user feedback.

In March 2025, Manus began offering broader access with free daily tasks for all users and a one-time bonus of credits. Mobile applications for iOS and Android were launched in late March 2025.

Subscription tiers

Manus offers multiple subscription options:

Free Tier: One free daily task (equivalent to 300 credits) with a one-time bonus of 1,000 credits for all users

Manus Starter: US$39 per month including 3,900 credits per month and ability to run up to two tasks concurrently

Manus Pro: US$199 per month including 19,900 credits per month, ability to run up to five tasks concurrently, and access to beta features

Manus Team: US$39 per seat per month with a five-seat minimum (US$195/month total), providing 19,500 shared credits and dedicated infrastructure

Both paid plans provide enhanced stability with dedicated resources, extended context length, and priority access during peak hours.

Reception

Industry response

The launch generated significant attention within the AI research community and technology industry. Technology journalist Rowan Cheung, founder of The Rundown AI newsletter, described Manus's launch as "China's second DeepSeek moment" and called it "REALLY good," comparing it to a combination of OpenAI's Deep Research, Operator, and Claude Computer. Former Google employee and AI-focused YouTuber Bilawal Sidhu called Manus "the closest thing I have seen to an autonomous AI agent."

Jiang Chen of MIT Technology Review described using Manus as "like collaborating with a highly intelligent and efficient intern," though noting that Manus sometimes lacked understanding of objectives, made incorrect assumptions, and cut corners. Jack Dorsey, co-founder of Twitter, praised Manus as "excellent."

Some experts expressed skepticism about claimed benchmark results. Kyle Wiggers of TechCrunch reported that Manus had trouble with seemingly simple tasks such as ordering a sandwich, booking a hotel room, or developing a game, noting it "didn't work quite as well as advertised." Users also reported error messages, endless loops, factual inaccuracies, and crashes during the early beta period.

Media coverage

Forbes described Manus as potentially changing the notion that the United States is the undisputed leader in AI development, suggesting China may have surpassed the U.S. in the race to fully autonomous AI systems. The platform was featured in Chinese state media broadcasts, signaling government support for the technology.

Comparison with competitors

Manus competes in the emerging autonomous AI agent market alongside:

• OpenAI Deep Research – Research-focused agent integrated with ChatGPT
• Google Project Astra – Multimodal AI assistant with agent capabilities
• Anthropic Computer Use – Claude-based system for computer control
• Microsoft Copilot Studio – Enterprise-focused automation platform
• Devin – AI software engineer by Cognition AI

Manus differentiates through its consumer-focused approach, transparent execution interface, and multi-model architecture strategy. Unlike traditional chatbots that require step-by-step instructions, Manus is designed to start tasks independently and dynamically adjust its approach.

Controversies and concerns

Safety and control

AI safety researchers have raised concerns about autonomous agents operating with minimal human oversight. Questions include potential for unintended actions with real-world consequences, difficulty in predicting agent behavior in novel situations, challenges in establishing meaningful alignment with user intentions, and risk of agents accessing sensitive information inappropriately.

The company is aware of relatively high failure rates compared to systems like ChatGPT, with users reporting crashes and a tendency to get stuck in infinite feedback loops. Chief Scientist Ji Yichao acknowledged these as teething issues inherent in launching new autonomous agent technology.

Model transparency

Critics have noted limited transparency regarding which foundation models handle specific tasks and how the system routes requests. This raises questions about reproducibility of results, attribution of capabilities to specific models, verification of benchmark performance claims, and understanding of system limitations.

The company describes Manus as operating through "extreme repackaging" or "kitbashing" of existing large language models rather than developing proprietary models from scratch. Some industry observers question whether this approach represents genuine innovation or merely clever integration of existing technologies.

Regulatory scrutiny

Benchmark's $75 million investment in Manus came under review by the U.S. Treasury Department in May 2025 over compliance with 2023 executive order restrictions on U.S. investments in Chinese AI companies. Benchmark's lawyers cleared the investment on grounds that Manus does not technically develop its own AI models and is incorporated in the Cayman Islands rather than mainland China. The investment attracted criticism from Founders Fund partner Delian Asparouhov.

Future development

Planned features

According to company statements, Butterfly Effect plans to expand Manus availability to additional regions and languages throughout 2025, with particular focus on markets including the United States, Japan, and the Middle East

External links

• Official website