Manus

Manus is an autonomous general AI agent built by Butterfly Effect, a startup founded by Xiao Hong that launched March 6, 2025 and went viral within 72 hours on Chinese social media. Manus claims state-of-the-art performance on the GAIA benchmark across all three levels — 86.5 percent on Level 1, 70.1 percent on Level 2, and 57.7 percent on Level 3 — beating OpenAI Deep Research at 74.3, 69.1 and 47.6 percent. Pricing starts free with 300 daily refresh credits, then 20 dollars per month for the Standard tier, 40 dollars per month for Customisable, and 200 dollars per month for Extended. Score: 7.8 out of 10.

What Is Manus?

Manus is an autonomous AI agent — not a chatbot, not an assistant, not a copilot. It is a system that takes a single-sentence brief and drives a complete task to completion: research, plan, execute, deliver an artifact. It was built by Butterfly Effect Pte. Ltd., a company founded by Chinese entrepreneur Xiao Hong, and operates under the Monica.im umbrella brand. The company is now Singapore-registered but was originally headquartered in Beijing, a detail that matters later in this review.

Manus launched on March 6, 2025. Within 72 hours, the invite-only access waitlist hit six figures and the product partner publicly admitted they had "underestimated the enthusiasm of the public response." Western press called it a "DeepSeek moment for AI agents." By April 2025, Benchmark led a 75 million dollars Series B at a 500 million dollars valuation with Tencent and HSG (formerly Sequoia China) participating. By December 2025, Meta announced an acquisition reportedly worth 2 to 3 billion dollars. Manus 1.5 shipped October 16, 2025. The desktop app shipped March 16, 2026.

What differentiates Manus from OpenAI Operator, Claude Computer Use, Cognition Devin, Replit Agent and Google Gemini Deep Research is a single architectural bet: instead of one model driving a rigid tool schema, Manus runs multiple specialised agents in an isolated Linux sandbox and lets the executor write Python on the fly to call any library as an action. The paradigm is called CodeAct. It means Manus can do things the others structurally cannot — like combining pandas, playwright and a scheduled cron in one task without a human wiring the tools together.

What Manus Actually Does

Understanding Manus requires dropping the "ChatGPT but more autonomous" mental model. A Manus session is closer to hiring a junior analyst with a laptop, coffee and 24 hours.

The agent loop

Every Manus session runs a structured loop: analyse the event, select the tool, execute the command, observe the result, decide the next step. The loop iterates until the task is done or the credit ceiling is hit. The user sees a live timeline of actions — "searching flight APIs for April 20 to 25 Tokyo," "scraping tourism board for Ghibli Museum availability," "composing itinerary.pdf" — and can interrupt or redirect at any point.

Multi-agent roles

Inside the session there are three specialised agents coordinated by the Executor:

• Executor Agent — the front-end interface, owns task orchestration, talks to the user. Does not know the internals of the sub-agents.
• Planner Agent — decomposes the brief into a plan, writes it to a scratchpad, updates it as reality hits.
• Knowledge Agent — handles information retrieval, multi-source synthesis and fact-checking.

Manus engineers published their approach in a blog post titled "Context Engineering for AI Agents: Lessons from Building Manus." The short version: context quality and persistence across sub-agents drives quality more than model choice.

The sandbox

Each session spins up an isolated Linux environment. Inside are a full shell, a chromium browser via the open-source Browser-use framework, the Python interpreter with the pip ecosystem, a working directory for file I/O and a publishing endpoint that can deploy a website to a manus-subdomain.com URL in one step.

CodeAct — Python as the universal tool

Most agent frameworks (LangChain, ReAct-style systems, Operator) emit structured tool calls — typed JSON describing an action. Manus follows the CodeAct paradigm from a 2024 research paper: the agent's action is a Python snippet executed in the sandbox. The practical consequence: Manus can chain logic across pandas, BeautifulSoup, matplotlib, playwright and a REST call in one step, and it can handle conditional branches naturally. The cost: debugging a failing action requires reading Python, not a clean error schema.

GAIA Benchmark Claims — SOTA Across All Three Levels

Manus's core marketing claim is GAIA SOTA. GAIA (General AI Assistants) is a benchmark published by Meta, HuggingFace and AutoGPT researchers that tests an agent on real-world, tool-use heavy tasks — the kind of multi-step problems where you actually need to browse, compute and reason. It has three levels by difficulty.

The largest delta is Level 3 — the multi-step, tool-heavy tasks — where Manus leads by roughly 10 percentage points. That is where autonomous agent quality actually shows up in practice. The usual disclaimers apply: these numbers are self-reported by Butterfly Effect on a benchmark Manus clearly optimised for, and independent third-party replication has been thin. But on our own task testing (see the "our experience" section below) the agent clearly runs a longer horizon than Operator or Claude Computer Use before hallucinating or getting stuck.

Manus Pricing (2026)

Manus uses a credit-based model with four paid tiers plus a Team plan. Credits fund every action — a search, a scrape, a Python execution, a slide generation. More complex or longer-running tasks consume more credits. A simple itinerary might cost 300 credits. A wide-research dossier can burn 4,000 to 10,000 in a single run.

Annual billing saves roughly 17 percent across paid plans. The Team plan unlocks SSO (SAML, OIDC), shared credit pools, usage analytics and shared templates — the features enterprise buyers actually need. There is no published Enterprise tier with SLA as of April 2026; large deals go through direct sales.

The credit trap: the 20 dollars per month Standard plan sounds generous at 4,000 monthly credits plus 9,000 refresh credits (300 per day times 30). In practice, two Wide Research runs in a day can consume the entire daily refresh, and three complex coding sessions in a month can exhaust the 4,000 reserve. Serious users land on the 40 dollars per month or 200 dollars per month tiers within the first billing cycle. Plan accordingly.

Manus vs OpenAI Operator

OpenAI Operator is a browser-use research preview optimised for e-commerce, form filling and cross-site transactions. Manus is a general-purpose agent that can browse, but also executes Python, reads files and deploys sites.

• Horizon: Operator excels at short, clearly scoped browser tasks (book a flight, fill a form). Manus is stronger on multi-hour, multi-source tasks that mix web and code.
• Artifacts: Operator returns a browser state and a message. Manus returns files — PDF, XLSX, PPTX, a deployed website.
• Compliance: OpenAI has SOC 2 Type II, ISO 27001 and a mature enterprise posture. Manus does not publish equivalent certifications as of April 2026.
• Availability: Operator requires ChatGPT Pro at 200 dollars per month. Manus has a free tier and a 20 dollars per month entry point.

Manus vs Claude Computer Use and Claude Agent SDK

Anthropic's Claude Computer Use (released late 2024, evolved into Claude Agent SDK by mid-2025) lets Claude operate a virtual desktop — move the cursor, click, type. It is a primitive capability that developers wire into their own agent loop.

• Abstraction level: Manus is a product you use. Claude Agent SDK is a primitive you build on.
• Model: Claude Computer Use runs Claude. Manus routes to the best model per task (historically Claude Sonnet and DeepSeek).
• Sandbox: Claude Computer Use controls a real desktop. Manus controls a headless Linux environment with a browser inside it.
• For whom: Claude Agent SDK if you are a developer shipping your own agent. Manus if you want the agent already built.

Manus vs Cognition Devin

Devin is Cognition's autonomous software engineer, positioned as a coworker that takes tickets and opens PRs. Manus is broader — it will write code, but also plan a trip, screen resumes, analyse stocks.

• Scope: Devin is a vertical agent for software engineering. Manus is horizontal general-purpose.
• Reliability: Devin's public success rate on SWE-bench-style tasks has been around 13.86 percent, well below its marketing. Manus does not chase SWE-bench but delivers more consistent completions on the GAIA-style generalist distribution.
• Pricing: Devin is 500 dollars per month on the entry Enterprise tier. Manus starts free.

Manus vs Google Gemini Deep Research

Google's Gemini Deep Research is a research-specific agent bundled with Gemini Advanced. It is very good at what it does — synthesising sources into a long-form report.

• Breadth: Deep Research is a research tool. Manus is a research tool plus a coding tool plus a deployment tool plus a scheduler.
• Output: Deep Research outputs a well-cited report. Manus can output the report, the spreadsheet, the deck and a landing page.
• Governance: Google Workspace compliance is a known quantity. Manus is not.

Manus vs Replit Agent

Replit Agent is a code-first autonomous agent that ships prototypes to Replit's hosting. Manus will also ship a site, but the hosting is a sub-feature rather than the whole product.

• Infrastructure: Replit Agent assumes Replit as the stack. Manus is stack-agnostic.
• Use case: Replit Agent for coders who already live on Replit. Manus for anyone who wants a generalist agent.

Data Residency and Privacy — the China Question

This is the section you cannot skip if you are evaluating Manus for anything sensitive. The core facts:

• Butterfly Effect is registered in Singapore. The founding team, early engineering team and significant infrastructure footprint originated in Beijing.
• Independent security researchers have traced outbound Manus traffic to servers in Shenzhen, China. Butterfly Effect has not published a definitive data-residency statement.
• China's National Intelligence Law (Article 7) requires organisations and citizens to "support, assist and cooperate" with state intelligence work. Any data flowing through PRC-controlled infrastructure is legally compellable.
• In March 2026, Chinese authorities imposed exit bans on Manus executives pending a review of whether technology developed in China fell under national security or export control regulations.
• Meta's December 2025 acquisition (reportedly 2 to 3 billion dollars) has Manus operating as a stand-alone brand from Singapore, with Meta stating the Manus service would continue. The regulatory review complicates that narrative.

Practical implications for your buying decision:

• Regulated industries (healthcare, finance, legal, defence, government): do not use Manus on client data until Butterfly Effect ships SOC 2 Type II, ISO 27001 and an EU or US-only data-residency commitment in writing.
• Personal and marketing use cases: the risk profile is lower but not zero. Assume anything you type in Manus is potentially visible to a Chinese state actor via legal compulsion.
• Enterprise EU buyers: GDPR cross-border transfer obligations (Schrems II) apply. Without an explicit transfer mechanism, using Manus on EU personal data is a compliance risk.

None of this is unique to Manus — every Chinese-origin AI tool (DeepSeek, Qwen, Kimi) faces the same scrutiny. The difference is that Manus is an autonomous agent with access to your files, credentials and browser sessions. The attack surface is broader than a chatbot.

Real Use Cases We Tested

1. Competitive research

Brief: "Map the top 10 autonomous AI agents as of April 2026, build a comparison spreadsheet with pricing, GAIA scores, and compliance posture, and output a slide deck." Manus delivered in roughly 40 minutes, consumed around 3,800 credits. The deck was 12 slides, the sheet was accurate on 8 of 10 tools, two pricing rows were outdated by one version.

2. Coding

Brief: "Build a small Next.js dashboard that fetches the top 10 crypto coins from CoinGecko, shows a sparkline per coin, and deploys to a public URL." Manus shipped in 28 minutes, one deployed URL, one minor CSS defect. Credits consumed: around 2,200.

3. Scheduled automation

Brief: "Every Monday at 9am Bali time, scrape the top 5 AI tool launches on Product Hunt, summarise them in a table, and email me the result." Manus set up the scheduled task cleanly. It has been running for 6 weeks at the time of this review with one skip (Product Hunt rate-limit, handled gracefully the following week).

4. Stock and data analysis

Brief: "Pull Tesla stock data for the last 12 months, chart the 50-day and 200-day moving averages, write a one-pager on what happened." Manus delivered the chart, an XLSX with the raw data and a two-paragraph narrative. Around 1,400 credits.

5. Resume triage

Brief: "Here are 15 resumes for a senior ML engineer role. Rank them against this JD and explain." Manus produced a ranked table with rationale per candidate. Two ranks we disagreed with on expert review. Around 900 credits.

Pros and Cons

Pros

• True end-to-end autonomy — ship briefs, get artifacts.
• GAIA SOTA claim with a 10-point Level 3 lead over OpenAI Deep Research.
• Multi-agent + sandbox + CodeAct architecture is genuinely differentiated.
• Free tier is the most generous in the autonomous agent category.
• Desktop app reduces cloud-only dependence.
• Scheduled tasks are production-useful out of the box.

Cons

• China origin plus Shenzhen data flows plus National Intelligence Law = real risk for regulated sectors.
• Credits drain fast — Wide Research can cost 4,000 credits per run.
• Reliability is not Anthropic or OpenAI-grade yet — sessions can hang on bot walls.
• No SOC 2, ISO 27001 or HIPAA listings published as of April 2026.
• Ownership churn — Meta acquisition plus Chinese exit bans create uncertainty for long-horizon procurement.
• Support is a ticketing system on Free and Standard tiers.

Our Experience

We have run Manus for roughly 40 hours of wall-clock sandbox time across research, coding, analysis and automation tasks. The headline impression: when Manus works, it works at a qualitatively higher level of autonomy than Operator or Claude Computer Use — it will keep going for 30 to 60 minutes, pivot around obstacles, and deliver a complete artifact. When it fails, it fails opaquely — the session hangs, the credits keep burning, and the only recourse is to kill the task and restart with tighter scope.

The sweet spot is well-scoped but multi-step briefs that involve 5 to 15 tools and a clear final artifact. The anti-pattern is open-ended exploration — "figure out our go-to-market" will consume credits without converging. The anti-pattern is also anything behind aggressive bot detection — Cloudflare challenge pages, LinkedIn login walls and modern CAPTCHAs will block a Manus session hard.

On latency: a typical Wide Research task runs 15 to 45 minutes. A coding-plus-deploy task runs 20 to 60 minutes. That is too slow to replace a chatbot. It is the right speed to replace a junior analyst for a bounded task.

Who Should Use Manus?

Ideal users

• Solo founders and small teams who want a generalist agent for non-sensitive research, analysis and content tasks.
• Marketers running competitive scans, launch pages and one-shot campaigns.
• Indie developers who want a research-plus-build-plus-deploy loop without assembling the agent stack themselves.
• Analysts doing due diligence, market mapping or multi-source literature review on public information.
• Ops leads who want scheduled automation without writing the cron and the scraper themselves.

Not the right fit for

• Regulated industries (healthcare, finance, legal, government) handling client or patient data.
• EU enterprises bound by GDPR cross-border transfer rules without a legal review of Manus's data flows.
• Teams that need 99.9 percent SLA, dedicated support and published compliance certifications.
• Developers already productive with Claude Agent SDK or OpenAI Assistants — the DIY path is more flexible if you already have it.

Verdict: 7.8 out of 10

Manus earns a 7.8 out of 10. It is the most architecturally ambitious autonomous agent on the market as of April 2026, and its GAIA numbers are real enough to matter. The multi-agent sandbox plus CodeAct plus scheduled tasks combination is a genuine product advantage that OpenAI Operator, Claude Computer Use and Google Gemini Deep Research do not match on autonomy horizon.

Points are lost for three reasons: the China origin plus traced Shenzhen data flows plus National Intelligence Law exposure is a non-trivial procurement risk; the credit economy pushes serious users to 40 dollars per month or 200 dollars per month within weeks; and the compliance posture is not yet Fortune 500-ready. The Meta acquisition and Chinese exit-ban review add further uncertainty.

Score breakdown:

• Features (9.0 out of 10): Multi-agent + Linux sandbox + CodeAct + scheduled tasks + desktop app is the broadest autonomous feature set shipping today.
• Ease of use (7.5 out of 10): Great UX when it works, opaque failures burn credits.
• Value (7.5 out of 10): Free tier is excellent, Standard is tight, Extended is fair for heavy users, Team pricing is competitive.
• Support (7.2 out of 10): Docs are improving, ticketing only on lower tiers, no published SLA.

Frequently Asked Questions