Manus
General AI agent from Butterfly Effect, the Wuhan and Beijing team now based in Singapore. Runs long multi-step tasks in a cloud VM. Blocked from mainland China since 2025.
Overview
Butterfly Effect Pte. Ltd. was founded by Xiao Hong in 2022, worked out of Wuhan and Beijing, and moved its headquarters to Singapore in mid 2025. Manus launched in beta on 6 March 2025 and did something no Western lab had shipped as a product: it gave a language model an isolated cloud virtual machine with a shell, a browser and a file system, then let it work unattended for hours. Chief scientist Ji Yichao and product lead Zhang Tao built the loop around third party frontier models rather than training their own, which is worth stating plainly because it shapes everything about the product's economics and its geography.
The current release is Manus 1.6, dated 15 December 2025. It split the product into three model tiers, added a fast Chat Mode for short questions and introduced Wide Research, which spawns dozens of sub-agents in parallel across sources rather than walking one long chain of thought. That is the feature worth understanding technically. Breadth-heavy tasks like comparing 100 items or generating 50 design variants map badly onto a single sequential agent, and parallel fan-out is a genuine structural fix rather than a bigger context window. Scheduled tasks and external connectors followed through May 2026. Pricing is credit-metered: free users get 300 daily refresh credits, 20 USD buys 4,000 monthly credits and 40 USD buys 8,000, with a 200 USD tier above that.
The China practicalities are blunt. Manus is not available in mainland China, the Chinese-language social accounts were closed and the planned domestic version was shelved, because the agent depends on models that are not legally servable inside the mainland. From Shenzhen you need an overseas account and a network route to reach it. Then there is the corporate story: Meta announced an acquisition in December 2025 at a reported two to three billion dollars, China's National Development and Reform Commission blocked the deal on 27 April 2026 after a Ministry of Commerce review that included exit bans on executives, Meta formally cut ties on 15 June and Manus resumed operating independently on 11 August 2026. The product survived and is still shipping. Whether you want a year of that behind a tool in your stack is a separate question.
Key Features
- ✓ Agent mode executes inside an isolated cloud VM with shell, browser and file system access, so a task keeps running after you close the tab
- ✓ Wide Research fans one task out across dozens of parallel sub-agents, shipped with Manus 1.6 on 15 December 2025 for breadth-heavy jobs like ranking 100 programmes or screening 1,000 stocks
- ✓ Three model tiers inside one product: Manus 1.6 Lite, 1.6 and 1.6 Max, chosen per task rather than per account
- ✓ Scheduled tasks and connector integrations added through May 2026, so recurring jobs run against external services on a timer
- ✓ Chat Mode added in 1.6 as a fast path for short questions, kept separate from the slow agent loop
- ✓ No in-house foundation model. Manus orchestrates third party frontier models behind its own planning and execution layer
Where it holds
- • The clearest proof that a Chinese team could ship a general agent product ahead of the American labs. In March 2025 it was first, and the category took its current shape partly because of this demo
- • Cloud VM execution means a two-hour job survives a closed laptop, which most agent wrappers still get wrong
- • Wide Research is a real architectural answer to breadth-heavy work, not a longer context window with better marketing
- • English interface and documentation are first class. This was built for a global audience from early on, which is rare for a product with these origins
Where it breaks
- • Not usable from mainland China. The company blocked mainland access and closed its Chinese social accounts in mid 2025, so from Shenzhen you need a non-mainland account and network path
- • Credit costs are hard to forecast. A heavy run can burn several hundred credits out of the 4,000 you get on the 20 USD plan, unused credits do not roll over and nothing warns you beforehand
- • Closed all the way down. No weights, no self-hosting, no local model option and no visibility into which frontier models sit behind the agent loop
- • Ownership chaos through 2026. Meta announced an acquisition in December 2025, China's NDRC blocked it on 27 April 2026, Meta cut ties on 15 June and the company returned to independence on 11 August 2026. That is a lot of instability for a vendor you plan to build on
My Take
Being first still counts for something. In March 2025 this was the product that made "AI agent" mean a cloud VM grinding through hours of work rather than a chatbot with a tool schema, and the Wide Research fan-out that shipped with 1.6 on 15 December 2025 remains a sharper answer to breadth-heavy tasks than most rivals manage. The sting from inside China is that you cannot use it here at all: mainland access was cut in 2025 and the team sits in Singapore now, which tells you exactly who the customer is. Budget carefully, because one deep run can eat several hundred of the 4,000 monthly credits on the 20 dollar plan and nothing warns you first.
Quick Info
- Pricing:
- freemium
- Openness:
- Proprietary
- Starting at:
- Free tier: 300 daily refresh credits that reset every 24 hours, plus a one-time 1,000 credit signup bonus, 1 concurrent task and Manus 1.6 Lite only. Pro at 20 USD per month for 4,000 monthly credits, 40 USD per month for 8,000, and a 200 USD per month tier for extended usage. Team from 20 USD per seat per month, two-seat minimum. Annual billing saves 17 percent. Monthly credits do not roll over and a single complex run can consume several hundred, with no cost estimate shown before the run starts.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Manus holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Manus
OpenManus
open-sourceMIT-licensed open agent framework from the MetaGPT team, prototyped in three hours as the open answer to Manus. 58,100 stars and still pushing commits in August 2026.
OpenClaw
freeOpen-source AI agent framework for building autonomous AI systems that can interact with tools, APIs, and environments.
Devin
freemiumCognition's autonomous coding agent, now bundled with the desktop IDE that used to be Windsurf, aimed at teams offloading well-specified tickets to background runs.