Quick answer. Manus is a Singapore-based, Chinese-founded autonomous AI agent that crossed $100M ARR in December 2025 and was promptly bought by Meta for roughly $2B. China's state planner blocked the deal on April 27, 2026, ordering both sides to unwind it. The product is still live and shipping: in March 2026 Manus launched a Desktop app with a feature called My Computer that lets the agent execute terminal commands, read local files, and use installed dev tools on macOS and Windows.
Manus has been one of the loudest names in autonomous-agent land for the better part of a year, and 2026 has only turned the volume up. Between an unsolicited $2B Meta acquisition, a desktop app that pushes the agent off the cloud and onto your laptop, and a Chinese regulator that stepped in and said "no" to the deal, there's a lot to track. This is a current-state read for engineering teams trying to decide whether Manus belongs in their toolbelt.
What is Manus AI and what does it actually do?
Manus is a general-purpose autonomous agent built by Butterfly Effect, a startup that was founded in China and relocated its headquarters to Singapore in 2025. The product launched its first general AI agent in March 2025 and was described by Chinese state media as the "next DeepSeek" almost immediately.
Butterfly Effect was founded in 2022 by Yichao "Peak" Ji (chief scientist, MIT Tech Review Innovators Under 35), along with co-founders Xiao Hong and Zhang Tao. The company is headquartered in Singapore, with the founding team and early engineering rooted in China.
The shape of the product is straightforward: you give Manus a goal in natural language, and it plans, decomposes, and executes the steps to deliver a finished artifact. The agent runs in a virtual sandbox where it can browse the web, scrape data, fill forms, run code, edit spreadsheets, build slide decks, and write reports. Typical use cases the team highlights:
- Market research with cited sources
- Multi-step coding and small app generation
- Data analysis pipelines from raw CSV to chart and write-up
- Long-running browsing tasks (lead lists, comparison tables)
For a quantitative anchor: Manus 1.5 reportedly scores 86.5% on GAIA Level 1, 70.1% on Level 2, and 57.7% on Level 3 — the General AI Assistants benchmark that grades agents on real-world multi-step tasks. Those Level 2 and Level 3 numbers put Manus at the front of the consumer-facing general-agent pack as of mid-2026.
What model powers Manus?
Manus orchestrates multiple LLMs rather than betting on a single base model, but the substrate has been publicly reported as primarily Claude. The original Manus shipped on Claude 3.5 Sonnet; subsequent versions have tracked the Claude family forward through Sonnet 4 and into the Opus 4.x line, with a fine-tuned Qwen handling certain sub-tasks where a smaller, cheaper model is sufficient. AnalyticsIndiaMag literally titled a piece "Manus is a Wrapper of Anthropic's Claude, and It's Okay" — and that's a reasonable read of what's actually inside. The pitch versus a chat-first product like ChatGPT is that Manus was designed as an agent from day one, not a chatbot with tool-use bolted on top; the model layer is a means, not the differentiator.
Why did Manus blow up so quickly?
Two reasons: the demos landed at the right moment, and the revenue followed. By December 2025 — eight months after launch — the company reportedly crossed $100M in annualized revenue. That's an unusual ramp even for the current cycle, and it was the trigger for the acquisition conversation with Meta.
The product also got luck on positioning. While OpenAI's Operator and ChatGPT Agent Mode were still framed as research previews for many users, Manus had a polished consumer surface, a credit-based pricing page, and a stream of viral threads showing the agent autonomously planning travel, running competitive research, and shipping working code without a human in the loop. The autonomy-vs-supervision gap was wider than the actual capability gap, but in the perception market that gap matters.
What happened with Meta's $2B acquisition?
Meta announced its acquisition of Manus in late December 2025 for a deal value reported between $2B and $3B. Two things happened immediately after the announcement:
- Meta started integrating Manus into its internal systems, and a number of Manus executives joined Meta.
- In January 2026, China's Ministry of Commerce opened an assessment of whether the transaction complied with export-control, technology-transfer, and overseas-investment rules.
That investigation ran for roughly three months. On April 27, 2026, China's National Development and Reform Commission — the state planner — issued a one-line notice prohibiting foreign investment in Manus "in accordance with laws and regulations" and asked both companies to withdraw the transaction. Meta said in response that the deal "complied fully with applicable law" and that it anticipated "an appropriate resolution to the inquiry," without saying what that resolution looks like in practice.
The unwinding is genuinely messy. Reporting from CNBC and CNN notes that Manus had already been wired into Meta's stack and that staff had moved across. Reversing a half-completed merger four months after announcement is not a clean operation, and as of late May 2026 neither side has publicly disclosed how the integration is being separated, what happens to the Manus team that joined Meta, or whether the companies will pursue any restructured version of the deal.
Why did China block the deal?
The state planner did not elaborate beyond its short notice, so any account of motive is necessarily inference rather than fact. What the public reporting establishes is the context, not the intent:
- Manus was founded in China and only relocated to Singapore in 2025. Beijing has historically treated Chinese-rooted AI capability as strategic.
- The block came during a period of intensifying US-China tech competition, with export controls on advanced chips and ongoing disputes over outbound technology transfer.
- This is the second high-profile move by China to constrain a US acquirer of Chinese-linked AI assets in the same cycle.
Whether the block is read as an industrial-policy decision, a retaliation signal, or a one-off applied to this specific transaction is something only future cases will clarify. For US buyers eyeing Chinese-founded AI startups, the precedent is the point: even a Singapore HQ and a US-friendly executive team did not insulate this deal from a Beijing veto.
What is the Manus Desktop app and the "My Computer" feature?
In March 2026, Manus launched a Desktop application for macOS and Windows with a flagship feature called My Computer. The pitch: take the agent out of its cloud sandbox and let it operate directly against your machine.
Concretely, with the Desktop app installed Manus can:
- Read and write files in directories you've granted access to
- Execute terminal commands (subject to a permission gate)
- Use already-installed development environments — Python, Node.js, Swift, Xcode
- Drive locally installed applications
- Use the machine's idle GPU to run training or inference workloads
- Accept tasks remotely from a phone or another device, turning an always-on Mac mini into a persistent worker
The security model is a mandatory per-command approval gate. Every terminal command surfaces a prompt with two options: Allow Once for individual review, and Always Allow for trusted recurring workflows. That's a familiar pattern for anyone who's used Claude Code or other dev-focused agent tooling, and it's a meaningful step up from "the agent runs whatever it generates inside a remote container you can't see."
How does Manus compare to other autonomous agents?
The autonomous-agent space has gotten genuinely crowded in 2026. Here's how Manus sits relative to its closest comparables.
Manus vs OpenManus
OpenManus is the community-driven open-source clone built to mirror Manus's core capabilities. It's free, self-hostable, and exposes the orchestration layer so you can swap LLMs and customize tool integrations. The trade-off is the usual one: you bring your own LLM API keys, you maintain the stack, and the polish gap with the hosted Manus product is noticeable. For teams with strong infra muscle who want full control over data flow and tool surface, OpenManus is the obvious choice. For everyone else, the hosted product is faster to value.
Manus vs ChatGPT Agent Mode
ChatGPT Agent Mode merged what used to be Operator (browser control) and Deep Research (multi-step report generation) into a single agent surface inside ChatGPT. It's tightly coupled to OpenAI's model stack and benefits from ChatGPT's existing distribution. Practically, Agent Mode leans more conversational — you and the agent stay in dialogue while it works — whereas Manus skews toward "set a goal, walk away, come back to a finished artifact." If your workflow is many short interactive tasks, Agent Mode fits. If it's a handful of long-running asynchronous jobs, Manus fits.
Manus vs Claude Agent SDK and Claude Skills
Anthropic's Claude Agent SDK and the related Skills system are aimed at developers building bespoke agents, not end users running a turnkey product. Claude Skills give you a way to package tool definitions, system prompts, and workflows that Claude can invoke. It's the closest thing to a building-block kit versus Manus's finished consumer/prosumer product. Plenty of teams will end up using both: Claude-powered Skills inside their own product, and Manus as an off-the-shelf research/automation surface.
Manus vs Devin
Cognition's Devin is the closest head-to-head most readers ask about, and the framing matters: Manus is a consumer-facing general agent — research, spreadsheets, slide decks, browsing, light coding — while Devin is a developer-focused SWE specialist built to take a ticket, branch a repo, run tests, and open a pull request. Devin lives inside an IDE-shaped workflow with git, terminals, and a code editor as first-class surfaces; Manus lives inside a chat-plus-files surface where code is just one of many output types.
The valuation gap reflects the positioning gap. Devin's parent Cognition has raised at a $10B+ valuation against a $400M war chest, betting that "AI software engineer" is its own category. Manus's pre-Meta raises were an order of magnitude smaller — and the proposed Meta acquisition was at roughly $2B — because the bet is breadth, not depth. On GAIA, Manus 1.5's reported numbers (86.5 / 70.1 / 57.7 across the three levels) put it ahead of Devin on general-agent tasks, but Devin remains the stronger pick for repo-aware work where merge-readiness, test-suite hygiene, and CI integration are the bar. See FutureAGI's Manus comparison for the deeper head-to-head.
Manus vs OpenAI Operator
Operator was the original browser-control demo from OpenAI and has been folded into Agent Mode for most consumer users. For pure browser automation tasks — buy this, fill that form, scrape this site — the two products are now broadly comparable on capability. Manus's edge with the Desktop app is that it isn't confined to the browser; it can drive the OS as well.
For a deeper read across the rest of the field — Devin, Cursor's Background Agent, Cline, Aider, and others — see our complete guide to AI coding agents in 2026.
Can teams use Manus today, and what does it cost?
Yes — the product is open to anyone, and the ownership uncertainty has not affected day-to-day availability. Manus uses a subscription-plus-credits model, where you pay a monthly base and the agent burns credits per task depending on length and complexity.
The current public tiers:
- Free — 300 daily refresh credits, 1 concurrent task, Manus 1.6 Lite in Agent Mode
- Standard / Pro — $20/month (or ~$17/month annual), 4,000 monthly credits
- Customizable — $40/month, 8,000 credits
- Extended — $200/month, 40,000 credits
- Team — $20/seat/month, 2-seat minimum
The credit model is the part to watch. Heavier tasks — long browsing sessions, large file generation, training-style workloads — burn through credits faster than chat-style turn counts would suggest. Teams running Manus seriously should budget at a tier above whatever the marketing copy implies for their use case, and re-evaluate monthly.
Where does Manus fit in an engineering team's stack?
For an engineering team in 2026, Manus is most useful as:
- An async research worker. Long-running competitive analysis, vendor evaluations, content audits, technical literature reviews — tasks where you don't want to babysit a chat window for an hour.
- A spreadsheet/document automation surface. The agent's web-plus-files combination is a good fit for "reconcile this CSV against this site" type jobs.
- A local-tools driver via the Desktop app. The My Computer feature is the most differentiated piece of the product right now. If you have an always-on machine and a steady stream of well-scoped local-file tasks, this is the surface to try.
Where it's still weaker:
- Deep code work in a real repo. For multi-file refactors, IDE-grade context, and PR-style workflows, dedicated coding agents (Claude Code, Cursor, Cline) remain stronger.
- Production automation. The credit model and the human-in-the-loop permission gate make Manus better suited to assisted work than to fully unattended cron-style automation.
- Sensitive-data workloads. The hosted product still does most of its work in the cloud sandbox; the Desktop app helps but doesn't fully resolve the data-egress question for regulated environments.
What should I watch next?
Three threads to keep an eye on through the rest of 2026:
- How the Meta deal actually unwinds. Reversed acquisitions of this size and integration depth are rare. The terms — what happens to Meta-employed Manus engineers, who retains IP, whether a restructured deal emerges — will matter for the next round of US-China AI M&A.
- Whether the Desktop app picks up real distribution. Local OS-driving agents are a category that's been promised for years and shipped only in pieces. If My Computer works well enough to be the default daily driver for a non-trivial number of users, that's a more significant product moment than the headline acquisition.
- Whether Manus pushes deeper into developer workflows. The team has so far positioned the product as a generalist. A serious move into code-agent territory — repository-aware editing, CI integration, IDE plugins — would put it head-to-head with Cursor and Claude Code rather than with ChatGPT.
FAQ
Is Manus still operating after the Meta deal was blocked?
Yes. The product is live, the website and apps are up, and subscriptions are processing normally. The block affects the corporate transaction, not the operating company.
Is Manus a Chinese company or a Singapore company?
Both, depending on how you count. The founding team and original engineering are Chinese; the company relocated its headquarters to Singapore in 2025 ahead of the Meta conversation. China's state planner still treated the transaction as one involving Chinese technology, which is why the deal was reviewable in Beijing in the first place.
What is the "My Computer" feature exactly?
It is the headline feature of the Manus Desktop app, released in March 2026. It lets the agent operate against your local machine — read and write files, run terminal commands, use installed dev tools, and drive applications — with a per-command permission gate.
How is Manus different from ChatGPT Agent Mode?
Manus was designed as an autonomous agent from day one and skews toward long-running asynchronous tasks where you set a goal and walk away. Agent Mode is built on top of ChatGPT's conversational surface and skews toward shorter, more interactive sessions. Capability overlap is significant; ergonomics differ.
What is OpenManus and should I use it instead?
OpenManus is an open-source community implementation of the Manus pattern. It's the right call if you need self-hosting, custom LLM routing, or full control over data flow, and you have the team to maintain it. Otherwise, the hosted Manus product is materially more polished.
Is Manus safe to give access to my local machine?
The Desktop app uses a mandatory per-command permission gate with "Allow Once" and "Always Allow" options, which is the right baseline pattern. As with any agent that can execute code locally, the practical safety depends on what you grant blanket permission to — be conservative with "Always Allow," especially for shell commands and filesystem writes outside a dedicated working directory.
What happens if the Meta deal is eventually restructured?
Unclear. Meta's public statement leaves the door open to "an appropriate resolution," which could mean anything from a minority investment to a licensing arrangement to a complete walk-away. Until either side says more, the public posture is that the acquisition is off and both companies are unwinding the integration.