Videos, podcast, websites, slides … easy to install, the tool offers a wide choice of possibilities. Even if he asks questions, in particular the management of data.
While the world has its eyes focused on American advances on AI, the Chinese appear as serious competitors for a few months. After the astonishment caused by Deepseek’s performances, it is Manus’ turn, created by a young Chinese, Xiao Hong, to cause surprise. Developed by the Chinese startup Monica, the AIA autonomous agent can perform different online tasks in an almost autonomous way, from a simple quick.
Manus works in an asynchronous cloud -based environment. The technical manus stack is limited. He calls on LLM, which could be Claude 3.5 Sonnet d’Anthropic and Qwen d’Alibaba. To operate, he interprets requests in natural language and generates action plans. LLMs are supplemented by deterministic scripts for data processing and system operations. When the user sends his prompt, manus analyzes the request, selects the tools and performs the commands. He can set out at each stage, until he succeeds in reaching the assigned task.
We wanted to test a certain number of its features, such as responding to a complex request, creating a website, an SEO audit, a video, a series of slides, a podcast and a video with comments. The opportunity to take charge of the tool and assess its capacities.
Fast handling
The installation of manus is very easy, a bit like that of Chatgpt. Just go to the website and create an account. The free mode offers 300 renewable credits every day, access to chat mode, the possibility of a simultaneous and planned stain. Note that credits are mainly consumed by LLM tokens, the use of virtual machines and third -party APIS.
We use the “Basic” offer, at 20 euros per month, for our tests. It includes in particular the use of advanced models in agent mode, the generation of images, videos or slides. We use “quality” mode for in -depth analyzes. It costs around 20 % more than the “speed” mode, which offers quick but less precise results. The tool also offers “plus” offers at 33 euros per month, “pro” at 166 euros per month and “team” at 39 euros per member per month.
1st test: a complex request
We make a complex request with the agent mode, to test the manusal capacities to respond to this style of question. Example of one of our prompts: “I am a tennis player who plays regularly. I have a rather offensive style. I am looking for a light racket of about 280 grams, balanced, for competition, at less than 200 euros. List me the main models available, with the advantages and disadvantages.”
To answer it, a bit on the model of the Nanobrowser extension associated with an LLM, the agent divides the task into different actions. He is looking for information on different websites, analysis and compiles. He notably creates Markdowns files, built on a light marking language. It is possible to accuse your request during research. After about 4 minutes, Manus generates a fairly relevant selection of the best rackets according to the criteria requested. Beware, however, of the verification of the responses provided.
2nd test: the creation of a website
Another possibility of manus, the creation of websites. No need for technical knowledge for this, a simple prompt is enough. The type of website, the main objective and the designer style of design. For example, we asked him to make a tennis racket comparison site, based on our previous prompt. In about ten minutes, the search tool for information, images and texts on the internet. During the process, he performs Ubuntu commands, instructions typed in the terminal to interact with the operating system. It reads HTML and CSS files, tests mobile responsibility, or builds a React project for deployment. At the end of its task, manus asks if the site should be kept locally or deployed publicly.
The result is surprising: the site is well structured, with clear sections. The UX is rather suitable for the time spent on it. The texts are relatively correct. It is possible to edit it directly with the button at the bottom right of the site. The company also claims that its AI can create “complex online stores with payments treatment, stock management and customer accounts.” To do with data security problems in particular.
3rd test: SEO audit
Among the many promises of manus, that of the SEO audit is particularly attractive. On paper, the possibilities of the tool can bring real progress on the subject. For our test, we asked him to make a SEO audit of a web page, by integrating the latest recommendations from Google and SEOs on the subject.
To answer it, we see that Manus seeks important documents, on the EEAT, the Core Web Vitals, or on the 2025 SEO trends. He performs JavaScript, or studies the structured data in JSON. At the end of his mission, three documents are generated, including a “complete SEO audit”, relatively successful from our point of view. It refers to recent concepts. Beware, however, of the veracity of words, such as advanced figures. Remember to check with the Google Search Console. The technical document and the recommendations are very generic and not fully exploitable as it is.
4th test: Creation of a scripted video
Manus has joined VEO3 recently. The Google IA video generator is available for Basic, Plus and Pro members. It should make it possible to create quality sequences, with a voiceover and sound effects.
Example of a prompt given: “Make a video of 40 seconds with these instructions: Western plan, two face-to-face men and robots behind. The two humans walk towards each other. Lighting: full day, sun which seems to bed in the background. The atmosphere is dominated by yellow, blue and brown. It is stifling. Salon, while the music of “Once upon a time in the west” of Ennio Morricone sounds.
We note that Manus uses in particular Ubuntu to generate the video. The result is interesting in terms of the quality of each scene. But the scenario is not well respected. We also do not recognize the characters from one scene to another. We do not have the requested soundtrack. This may explain the message displayed when preparing the video: “We cannot process this request due to the safety and compliance requirements of the community.”
5th test: generate a series of slides
Manus has also recently offered the creation of a series of structured slides from a simple prompt. This can be very interesting to use in the professional framework. For that,, It is possible to directly provide him with a file. Take care not to mention sensitive information. Thanks to the RAG, Manus can release a very interesting document.
For another of our tests, we rely on the knowledge taken from the Internet. The prompt: “I am a qualiopi trainer and consultant. Within the framework of the training that I dispense, creates a course based on clear and educational slides explaining the interest of passing the qualiopi certification. This must make it possible to easily understand the advantages of this one. The dominant colors of the slides are blue, yellow and red.”
In just a few minutes, Manus provides an usable document (see here). However, the information remains to be verified. We note the presence of logos, diagrams, graphics, or quotes that facilitate the reading of the document. It is well structured. The color distribution is questionable, it can be improved in the prompt.
6th test: Create a podcast
Manus can also create a podcast from a script. For example, we asked him to make a podcast from the article “Tutorial: how to automate Google Chrome with Nanobrowser and Gemini.” To create the requested audio file,, Manus analyzes the article, summarizes the structure and writes the script. It is possible to choose between a female or male voice to dictate the text.
The result generated is pretty good. The tone of the voice, the diction and the tone are quite convincing. Some downside: the pronunciation of certain words, in English in particular, leaves something to be desired, just like the verbose and technophile side of the generated text.
7th test: generate a video with comments
The creation of a video with a voiceover is also a possibility of manus. For example, we have submitted this prompt: “Make a video of this article” Tutorial: how to automate Google Chrome with Nanobrowser and Gemini “. She will take the form of a journalistic video taking up the large sections of the article. Illustrious-the with images related to the voice-over. These images must if possible change every 2 seconds. Use images and video extracts from the article.
The video generated has the same characteristics in terms of text as the podcast created. Unfortunately, the images are not always relevant with the text read, and their quality is not always optimal.
Watch out for shared data
Despite all these feats or potential, certain limits appear in Manus. If autonomy and latency are often satisfactory, the problem is rather from data security. Navigation agents who run remotely in the cloud can potentially have inappropriate behavior or leak data. Also, the handling of online platforms raises questions. Note that, according to our tests, Manus could not enter paid sites, even if the possibilities of the tool could undermine more than a paywall.
In addition, according to a recent study, elements from the manus’ confidentiality policy of March 8, 2025 only partially respect the general data protection regulations. For example, there is no mention of legal supervision of data transfers to Singapore, a third country not benefiting from a decision to adequate the European Commission. Data storage, up to 6 years, without clear purpose, also poses concern. The investigation also deplores the explicit absence of user’s consent for each purpose.
