How Manus went to market as a general AI agent with invite codes and X demos

The X angleManus demonstration videos spread on X after its March 2025 preview, and co-founder Yichao Ji used X to explain which models the agent used.

Manus arrived in March 2025 with a promise that no product had yet convincingly delivered to the public: a general AI agent that could take a single instruction and carry out a long, multi-step task on its own. The company behind it, known as Butterfly Effect and the team behind the Monica browser assistant, released an invite-only preview with an introductory video. Within days, Manus was one of the most discussed AI products on X. Its go to market strategy combined scarcity, shareable proof of work and unusually direct founder communication, and it reached a conclusion few expected within a year.

The product and the bet

Manus positioned itself as a general-purpose agent rather than a chatbot or a coding tool. Users described a goal, and the agent worked through it in a cloud sandbox: browsing websites, writing and running code, building files and assembling a final deliverable. Demonstrations showed tasks such as researching a market, building a spreadsheet model and producing a report from one prompt.

The demonstrations emphasized process as much as outcome. Rather than showing only a finished answer, the launch material showed the agent moving through a task: opening pages, writing code, building files. That made the agent feel less like a black box and more like a colleague whose screen you could look over.

Who it was for

At launch, the audience was effectively anyone curious about agents, filtered by who could get an invitation. The early adopters were AI enthusiasts, founders, analysts and creators, many of them active on X, who wanted to test whether autonomous agents had arrived. Over time, the product's natural fit became knowledge workers with research-heavy and document-heavy tasks: the kind of work that takes many steps across many tools but does not require deep domain judgment at each step.

Choosing such a broad audience runs against most advice on go to market strategy for startups. Manus could get away with it at launch because the novelty of the category did the segmenting. People who cared about agents found it.

The launch

On the evening of March 5, 2025, Butterfly Effect released an early preview of Manus, describing it as a general agent product. The introductory video was narrated by co-founder and chief scientist Yichao Ji, who also goes by Peak. The company also published claims that Manus outperformed other systems on the GAIA benchmark for general assistants, results that outside observers noted were the company's own.

Access was by invitation only. Reporting at the time said only a small number of people had received invitation codes. The company later explained the limits as a capacity constraint during an internal testing phase.

How X carried it

Manus is a textbook case of how scarcity and visible output interact on X.

The demo as the unit of distribution

The launch video gave X something to share on day one: an agent visibly doing multi-step work, which is far more striking in a feed than a chat transcript. As invitations trickled out, the audience shifted its attention from the company's footage to the question of whether the agent would hold up in the hands of others. In analysis, that is the moment when a gated launch either compounds or stalls, because third-party evidence carries more weight than anything a company films itself.

Scarcity as a story

Invite-only access became a story of its own. Business Insider reported that invite codes were being listed on resale sites, including the Chinese marketplace Xianyu and eBay, though it was unclear how many listings actually sold. Manus responded publicly that it had never set up paid channels for invitation codes and had not spent on marketing, and that it would release invitations gradually while prioritizing existing users.

In analysis, the resale market did two things for distribution. It signaled demand far more credibly than any company statement could, and it generated a second wave of coverage about the scarcity itself. It also created risk: fraud, frustration and a sense that access was arbitrary. Founders considering invite mechanics should weigh both sides.

A founder answering critics directly

The launch quickly drew skeptics. Some developers who examined the system argued that Manus was largely an orchestration layer around existing models rather than a fundamentally new model. On March 10, 2025, Ji addressed this on X, confirming that Manus used Anthropic's Claude 3.5 Sonnet along with fine-tuned versions of Alibaba's Qwen models, and that the team was testing newer Claude versions. He also explained aspects of the sandbox architecture.

This was an unusual choice. Many companies would have stayed vague about their model stack. Ji instead used X to describe the system plainly, which took some heat out of the critique and reframed Manus as an engineering achievement in orchestration and product design. The Decoder and other outlets reported his clarifications, extending their reach beyond his own followers.

Community beyond X

The momentum on X fed other channels. Reports said the official Manus Discord server grew very quickly in its first week, giving the company a place to handle support and feedback at scale while X remained the public stage.

What came next

Manus moved quickly from preview to business:

• Paid plans and mobile. On March 31, 2025, TechCrunch reported that Manus had launched two paid subscription plans and a mobile app. The paid tiers differed by credit allowance, the number of tasks that could run at the same time and priority access at busy times, so pricing was effectively a credit-based subscription.
• Opening the doors. In May 2025, Manus removed its waitlist and opened to all users, giving each user a small number of free credits per day plus a one-time bonus. Paid plans remained for heavier usage.
• Relocation and acquisition. The company later established its base in Singapore. On December 29, 2025, Meta announced it would acquire Manus. Reporting on the deal said Manus would continue to operate its subscription service, would keep its Singapore base and would discontinue operations in China, and that Meta planned to integrate the technology into its products, including Meta AI.

The sequence, from invite-only preview to paid plans within weeks to open access within about two months, shows a company using scarcity as a launch device rather than a long-term strategy.

The go to market lessons

Manus's path offers a compact set of lessons for founders launching agent products. These are analysis rather than company statements.

1. Make the process the demo

Showing an agent working, step by step, is more persuasive than showing only its output. For any agent startup, the shareable artifact should be the run, not just the result, and the product should make it easy for users to capture and share their own runs.

2. Scarcity amplifies, then frustrates

Invite codes created urgency and a secondary market that became news. They also created a backlash risk and a fairness problem. The analysis here suggests that scarcity works best when it is short, clearly tied to real capacity, and paired with a published plan to open up. Manus went from invites to open access within roughly two months, which is about the right horizon.

3. Answer the technical critique in public

Ji's X posts about the model stack show that candor can defuse a skeptical narrative. When experts are already reverse engineering a product, explaining the architecture costs little and earns credibility. A go to market team should prepare those explanations before launch, not after.

4. Monetize while attention is peaking

Paid plans arrived within a month of the preview. That captured willingness to pay while curiosity was highest and gave the company a revenue signal to set against its compute costs. A credit-based model also matched the variable cost of long agent runs, a pattern now common across saas go to market strategy for AI.

5. Benchmarks are claims, not proof

Manus led with benchmark results that it had run itself. Coverage quickly noted that. Founders should present self-reported benchmarks as such, and invite independent testing where possible, especially when launching into a crowded category.

6. Consider where the company is building

The move to Singapore and the subsequent acquisition by a US technology company point to a strategic consideration that sits beyond launch tactics: where a company is based affects which markets it can serve and which partners and acquirers it can work with. That is part of a full go to market strategy, even if it is rarely discussed at launch.

Scarcity, step by step

Manus moved through its launch phases unusually quickly. The table below summarizes the sequence as analysis.

A founder considering invite mechanics can treat this as a timing guide. The value of scarcity decays fast. The aim is to convert the attention it creates into paying users and an open product before frustration replaces curiosity.

Where it fits among go to market strategy examples

Manus belongs with Devin as an agent launched through video and gated access on X, and with DeepSeek as a product from a team with Chinese roots that captured global attention almost overnight. What sets it apart is the invite code frenzy and a founder who answered technical criticism directly. For more cases, see the examples library.