Manus AI Review

Research and a cited report

My first prompt asked it to compare five project management tools for small remote teams and return a cited feature matrix as a downloadable Word document. Within seconds it opened its own browser and began pulling live pricing straight off vendor sites.

Fifteen minutes later it handed back a finished .docx.

The report closed with a proper References section listing six dated sources, and a research note telling the reader to re-check pricing before signing anything. That last touch mattered to me, because pricing pages change and a static report goes stale the moment it is written.

This is the part Manus does well. It read the goal and did the digging itself, returning a file I could edit rather than a block of chat text. A normal chatbot would have given me a summary and left the sourcing to me.

Building an app from one sentence

The second test was harder on purpose. I asked Manus to build a real-time collaborative to-do app where several people could edit the same list at once, with login and a database behind it.

Before writing any code, it stopped to read one of its own internal skill files and run a pre-build check.

Then it worked, and kept working. My prompt went in at 17:39. Ten minutes later the system clock read 17:49 and the build was still on step three. Simple research finishes fast. This did not.

The moment that told me most came at the end. Manus finished the app and opened it in its preview pane, where it ran straight into the sign-in gate it had just built. Its sandbox browser could not clear the login, so it paused and offered me a choice: take control of the browser myself, or let it sign off on automated checks alone.

The app was not broken.

When I opened the same login in a separate browser tab, it signed me in on the first try. The block was inside Manus, not inside the software it produced.

Behind that gate sat a working product. CollabTodo had authentication, a live database, shared lists with a join code, and a presence counter showing who was active. It refreshed on its own every few seconds without a page reload. I added a task and watched it persist, then copied the share code to bring in a second person. For something built from one line of instruction in a single sitting, the feature set went well past a throwaway prototype.

So the coding result splits in two. Manus built more than I expected from one sentence of instruction. What it could not do was confirm its own work once that work sat behind a login, and that gap belongs to the agent rather than the code.

Manus AI, developed by Butterfly Effect Pte. Ltd., holds a 4.5-star rating on the Google Play Store. The listing records 4.58 lakh reviews against more than one crore downloads, and the page flags in-app purchases.

The rating sits high, but the written reviews split hard between praise and frustration.

A reading of the top reviews gives a sharper picture than the star average alone. Several users treat Manus as one of the most capable building tools they have used.

One reviewer called it the best AI “harness” they had tried. Their account was plain: you tell it what you want built and it starts hammering away in the background, pausing now and then to ask for input before it carries on. That user said they would hand out ten stars if the option existed.

Another five-star review focused on how approachable the tool feels. This user began with a simple Yes/No game idea and watched it grow into something interactive, calling the process of turning rough concepts into working projects effortless. They placed Manus among the AI tools they have enjoyed most.

The praise lands on one point. Manus shortens the distance between an idea and a project that runs.

The criticism runs the other way, and it is blunt. The sharpest complaints target the credit system rather than the core technology.

One user building an app with the Manus 1.6 Max model reported that the output came back full of bugs. Fixing a single bug ate 75 percent of their credits across several attempts, and the bug stayed broken. They also pointed out that no top-up option existed for buying more credits. Forty-four people marked this review helpful, the highest count in the set.

A second one-star review described a billing mismatch during a Pro trial. The user signed up for a seven-day free trial expecting the advertised 8,000 credits. By their account, the support reply stated that only 500 credits are granted during the trial window, with the full allocation arriving only once paid billing starts. Twenty-two people found the review helpful.

A three-star review sits between the two camps. This user called it a good app, then flagged a specific catch: credits are needed even to talk to the tool. Without them, a paying subscriber cannot check progress or see what has already been done.

One thread runs through the good and the bad. People respect what Manus can build, while the friction gathers around pricing and the gap between what a trial advertises and what it hands over.