The journey of how we built Manus

Hi everyone, I'm Tao, co-founder of Manus and right now acting as our chief product officer. Today I just want to

product officer. Today I just want to share you the journey of Manus. So on

this A16 top 50 AI apps list and you will see manus at the uh 31 base and also in this company uh we don't just have manus we have another product

called Monica and they both on the top D uh 50 apps and also if you are looking at another A16 they list which is where

startups spend their money on AI and you will also see that uh manus is also on this And uh just days ago, you know, like

five days ago, we just announced our first milestone, which is we achieved 100 million AR for Manas in less than nine months, which made us might be the

fastest company to achieve that from Z to 100 million AR. And not just that, we also like outperforming other frontier

model labs on some benchmarks just like this ROI benchmark from scale AI. And

you can see here on this benchmark uh minus 1.5 and one 1.0 Darl, they're at the first place and not just about the

percentage of completed tasks also if I are looking at the ELO score uh we are much higher than the second and and the

third place and this is just something we achieved at end of this year but I think most of you wonder like how did

all this get started. So today I just going to jump back to a much earlier time which is G when GBT3 launched.

Yeah. So and in 2021 OpenAI released their GBT3 model and uh but actually uh not many people noticed that including

us. So one year later which is 2022 and

us. So one year later which is 2022 and our CEO read he randomly just found the GBT3's model playground and saw its

potential that it can be a very good foundation model for efficiency tools and because red made some chrome plugins before so he decided that maybe he

should build a new chrome plugin based on GP3 model that is just one week before CH GBT and you know one we can later change it, den it. So the world

chang it and it makes no sense that we should build something else. So we were looking at if there's any like pinpoint

when people are using CH GB. So you know just like us like like this one yeah draft an email to PE which is our chief scientist about our journey at manus.

Yeah, a lot of people are using CH GBT, but a lot of people also do a lot of back and forth copy pasting before CH GBT and other working apps like this.

When you have the answer, you have to copy paste the subject to the email and

also you have to uh switch back and uh copy another content and back to the your working web page or working app.

Yeah, this is how sins works before and because we think that may be a painoint for a lot of users because right now there are a lot of working apps that

just uh inside the browser. So if we can develop a plug-in that can help our users to use AI inside the context without switching between different

tabs, I think that may be an opportunity for us. So uh like just one month after

for us. So uh like just one month after chip launched we released our first AI product which is modica. Modica is

something just a chrome browser plugin leaves in your browser. So this is modica. This sidebar when you click the

modica. This sidebar when you click the sidebar you will see there's a chatbot will jump out and you can just choose

any model you want in one product. You

can use Anthropia's model. You can use open eyes model and also definitely like Geminina. Yeah. Everything like deep

Geminina. Yeah. Everything like deep grog. Yeah. So you can just subscribe to

grog. Yeah. So you can just subscribe to Monica's membership but you can use a lot of different type of models. And

it's not just about a chat bot. It also

provides so many other features [snorts] which is just let me switch back to the Gmail scenario. Yeah. You you must

Gmail scenario. Yeah. You you must remember this is something we built nearly three years ago. At that time I remember Google doesn't have any like AI

AI capabilities for their product and at that time we've already have this function which is the Gmail writing assistant. So after you install our

assistant. So after you install our Chrome plugin and uh you can say that I don't want to type the prompt again. So

you can just uh tell Monica what do you want to write? Monica will write the email for you. And uh after we write the

email, you can just uh insert into your existing Gmail app. And uh we are not just stopping there. And if you are

watching a YouTube video, yeah, like this one. Yeah, you can also say after

this one. Yeah, you can also say after you install the Modica uh browser plugin, there will be a modica section that can help you transcribe the whole

video and make it into different sections. If you click to expand this

sections. If you click to expand this section, you will also see like uh what's a what's a timeline uh what the video is talking about. And if you click

on this, it will jump to the exact uh timeline. Yeah. With this time step.

timeline. Yeah. With this time step.

Yeah. So this is something we built like two years ago. Yeah. And uh also we have this like PDF tool. So if you if you are

reading a paper like this famous one, attention is all you need, you can just click the PDF summary and uh Modica will

just uh pass the whole PDF for you and uh you can just uh click on any paragraph to ask Modica to uh explain it or get some insight and also get the

whole summary of this PDF. It's very

convenient when you are reading like some papers. Yeah. And there's also like

some papers. Yeah. And there's also like another feature which is my favorite.

Yeah, a lot of AI products provides a AI summary feature but I I don't use them all because you know like when you are doing a AI summary, the AI doesn't know

your attention. So it doesn't know like

your attention. So it doesn't know like which part is important. So after you read the summary, you're still worried that you may miss something. So you have to read the original article again. That

happens to me. I don't know if you guys had the same problem. So I just uh designed this function. We call it the simplify article. So when you are

simplify article. So when you are reading some very long article and sometimes maybe very hard articles you can click when you after you install modica you can just click the simplify

article button and you'll see each paragraph will be much shorter and you can keep scrolling and you'll see along along the scrolling all these text will

be shorten nice and all the tables images they were just left there for you to understand the whole articles it will keep the original structure there and If

there are some paragraph you want to dig deeper, you can just move your mouse over here. The original paragraph is

over here. The original paragraph is still there. So it's very easy, you

still there. So it's very easy, you know, for you to read a very long article. Yeah. And uh this is just a

article. Yeah. And uh this is just a very small portion of what Monica can do because you know uh in the last three

years actually we made a lot of features into Monica. So if you are looking at

into Monica. So if you are looking at web looking at our website you will see we have like many apps including like image generation video generation uh PDF

tools uh and even we have desktop app, mobile app, a lot of things but the browser plug-in is the most frequent uh

usage you know for Monica's users. So

uh that is what happened to uh our first product Monica. Yeah. Uh it's kind of

product Monica. Yeah. Uh it's kind of like the best in its category. We have

like 3 million more users just for the Chrome browser plug-in and it's generating like 15 million AR even today. You know, even we have manus but

today. You know, even we have manus but Monica is still generating cash for us.

But we are not stopping there because you know like even mod is very successful as our first AI product but it's still a Chrome browser plugin and

as you know many users they don't know what exactly is a browser plugin in their whole lifetime they didn't install maybe just one plug-in which you know

makes the ceiling for us is very low. So

we are always thinking about what's our next step you know what's our next step.

So because you know like last year 2024 we were thinking about we are very good at making browser plugin so maybe we should build our own browser then we

don't have to ask our users to install plug-in because the browser is already there. So our first attempt of pirate is

there. So our first attempt of pirate is to pivot from a browser plugin to an AI native browser which something we never

release but I think today maybe the first time I show this product to the world because we never release it. Yeah.

This is a product we call it rift. Rift.

Yeah. And in this browser actually this is something we developed from last uh March you know like we spent like seven months from last March to last October

we spent seven months on this AI browser project and we we deal so many things inside it. As you can see here when you

inside it. As you can see here when you create a new tab uh we I personally call this feature all yeah which is can out

to group all the tabs you are browsing using a local run model. So your privacy just stays inside your browser and uh we

can just uh uh auto groupoup these tabs based on the web pages content. So you

don't have to group them by yourself.

Yeah, it's very convenient. And when you are uh browsing the web sometimes maybe you will see this image uh is not very

uh in high definition. So you can always click this HD scale to get this image into into HD scale. You will see the

difference here. Uh it's very

difference here. Uh it's very convenient. And also when you are

convenient. And also when you are reading this article, you can always ask the rift to summarize this whole section for you and you can just leave the card

at the side and when you are reading there will be some distractions at this link. uh you don't have you don't want

link. uh you don't have you don't want to be distracted. So you can always use the AI mode to click on this link summary. There will be a car and along

summary. There will be a car and along that you can keep scrolling reading other content and when the card is ready you can check okay what the link saying.

Yeah. After that you can just click keep keep scrolling. Yeah. And there are a

keep scrolling. Yeah. And there are a lot of other features like uh let the AI browser to extract some structured data

from some unstructured web pages. Yeah.

So actually rift is an AI later browser that have the entire modica built into this browser and has a local deployed

small size model uh running in the background and this is something we think it's very cool but after seven months of efforts we decided to sunset

this project. Why is that? Is because

this project. Why is that? Is because

you know along the way when we build this browser, we find it's very hard to convince someone switch from their

existing browser because everyone is so familiar with their current browser.

They don't want to switch to a a new one. Yeah. Even you tell me that oh this

one. Yeah. Even you tell me that oh this is something new. This is something very cool. This is something special. But

cool. This is something special. But

they are the the first thing they ask is not always about the AI features. It's

always about okay I have some features in Chrome. I have some feature in

in Chrome. I have some feature in Suffery. Do you have this? Do you have

Suffery. Do you have this? Do you have that? And as a very small startup we

that? And as a very small startup we can't meet all these requirements because you know like Chrome and Safy they're very big teams and they have a

lot of features we can't build in just the months. So after seven months of try

the months. So after seven months of try even we built so many AI features in this uh AI browser we decided that maybe this is not the right way to deliver the

next generation of AI experience on a old product. So we decided to sunset

old product. So we decided to sunset this project just uh I think just one week before the original release date.

Yeah. And there's a very dramatic moment because, you know, after we decided to sunset our AI browser project, we just

saw this video from the browser company uh because we really love Arc. Yeah. So

after we decide to sunset our uh AI browser rift, we also say that ARC is moving to DA and uh after years of of

trial, they also found out even they have the maybe they have the best uh new generation browser but they still can't convince their family members and the

friends to use Arc. That's something you know like we we can really resonate.

Yeah. So we think maybe now this is the time uh we really should end our AI little browser. Yeah. So that's our

little browser. Yeah. So that's our first attempt of of trying to build a new product that it failed after seven

months of trying and that is uh 2024 uh October and after we sunset the AI browser project we are thinking about okay now what's the next thing you know

what's the next thing and actually you know like from last year July uh something went super viral which is cursor uh a lot of a lot of people are using curs Because you know we guys we

are all coders engineers and we really love cursor. But the most interesting

love cursor. But the most interesting part about cursor is not about us engineers using it. It's about you know watching our friends our family members

who are long coders. They are starting to use cursor to solve their daily programs just like this. Uh this is an example for my wife. Yeah, my wife used

cursor to convert a video file to an audio file using cursor. Yeah. And uh

also uh my colleagues you know they their long coders they are using cursor to do some data visualization and calculations for their Excel. And when

we were watching these long coders using cursor, the first insight is they never care about the code editor side because they don't know how to evaluate the

quantity of the code cursor generated.

The only thing they care is the right panel of cursor and at that time you still have to click accept accept accept accept buttons. So the only thing they

accept buttons. So the only thing they do is keep pressing the accept button.

So that's kind of like give us the inspirations that maybe we should build something you know that is like the

right panel of cursor but combine the learning the knowledge we know from our AI browser project which is AI is super good at using browser but the problem is

we shouldn't let AI to use your browser because it kind of like the AI is competing with you for your own device.

It's like you hire an inter but you don't give thing a computer. Instead of

that you are asking the inter to use your computer. You are always competing

your computer. You are always competing with him. It's really weird. So

with him. It's really weird. So

combining the knowledge, the learning we get from our air browser project and the the inspirations we get from cursor, we decided that we should build the right

panel of cursor. But instead of running it locally on your computer, we give a virtual machine in the cloud for the AI

to use. So it's kind of like the right

to use. So it's kind of like the right kind of cursor but running in the cloud.

And it is not designed for engineers but for normal people for average users who who all the users that they don't know coding at all because we believe there

are already a lot of very fancy tools for engineers but we need to build something for average people. Yeah. So

that's kind of like the original idea of of manners. It's kind of like give the

of manners. It's kind of like give the LM a computer. Yeah. So after we decided

to uh build that so this is uh the basic concept of manus which is that you can just assign a task

to us and you'll see manus will have its plan. Yeah, this is manus's plan and

plan. Yeah, this is manus's plan and then manus will execute the plan step by step by itself and along these steps manage will do a lot of things in its

own computer which is like man is searching on Google and man is browsing these web page in its own browser and man is scrolling these pages man is

clicking through these different websites to find enough information yeah as as for this task you will see that I'm a gumpla which is Gandam model lover

So I want to build a collection side for all my models. Yeah. So man is writing all this code doing all this debugging

and finally at at the end after all this work man will deliver this very fancy website to me uh has all my uh all the

models images and you can market as your own collection and you can say uh this is my collections. Very cool. Yeah, just

use a very simple natural language to build that and you can see the idea behind that is just let the model to decide which is the next action to do

and then we perform these actions in the virtual machine and then we return the observations get after the action is performed in the virtual machine we return the observation back to the agent

architecture and let the agent to decide then what's the next action. So the

whole manus is just a very simple concept. Yeah, it's just a we something

concept. Yeah, it's just a we something we write uh at the bottom of our official website which is less structure more intelligence. We have zero

more intelligence. We have zero predefined workflows inside manus instead of that we just leave all the decision making to the foundational model to decide what's the next action.

So the whole manus agent architecture is just a lack action prediction machine.

Uh but definitely we have a lot of other secrets about like how to keep these things running and how to maintain this context. Months later, there's a new

context. Months later, there's a new buzzword which is context engineering, right? And our chief scientist pig uh he

right? And our chief scientist pig uh he wrote a very useful piece if you want to build an agent product about context

engineering and he also has another sharing with longchain this webinar uh like taining a lot of like insights and learning along the way we build manus.

If you are really want to build some uh advanced agents, I think these two pieces is something you must read and watch. Yeah. So this is kind of like the

watch. Yeah. So this is kind of like the basic concept of manners less structure more intelligence and uh uh it took us

like four months to development the first version of manus and in this year 2025 March 5th we decided that this is

the time we should launch the product to the world and just uh I think it's just like maybe six days before the launch when I looking at our launch website. I

saw it's kind of like empty. So, I

decided that maybe we should make a video about that, but there's only six days left. I call a lot of agencies and

days left. I call a lot of agencies and they all told us they need two weeks to uh to to to do the video production. So,

I decided that maybe I should shoot the video by myself. So, as you can see, our our original office is very small. So we

just use our that is a co-working space.

So we just use the public space of our co-working space to shoot this video.

And uh uh P yeah is our chief scientist.

He's the guy in the video and he's our inter uh helping me uh to shoot this launch video. Yeah. So it's my own

launch video. Yeah. So it's my own camera and uh I just use cap card uh to edit all the all the video. This is the

original pro project in my cap card.

It's still there. Yeah. So, this is the only money I paid for our launch video.

Yeah. Just for the membership of the cap card. So, this is the how we made this

card. So, this is the how we made this launch video. And uh at the in the

launch video. And uh at the in the launch uh we just hit the launch button together. Yeah. because it's a very

together. Yeah. because it's a very small office. We only have like uh like

small office. We only have like uh like six people in that office. Yeah. And we

celebrate the launch with our uh another office people. They are in this screen.

office people. They are in this screen.

You you'll see here. Yeah. We use a screen to see the remote office and we launch the manus together at the March 5th night. Yeah. We were very exciting

5th night. Yeah. We were very exciting that because you know we took so much so much belief in this product. We think

it's going to be a very successful product but we never expect that it will be this kind of success. You know just uh maybe like two days or three days

after our uh first launch we got super viral on social network on news media.

Yeah. This is the this is a photo that our CEO Red and I we were watching YouTube you know there there are a lot of people uh making videos about manas

and but you know like we expect it's going to be a very successful one but we never expected it would be this level of success so there are a lot of people

making video making videos and uh uh tweeting a lot of us uh like uh kam yeah from hugging face uh roam dby

even Jack you know like as you can say I use Twitter from 27 February so actually I use this platform uh for more than 18

years I never expect that someday the founder of Twitter will retweet my product it's really an honor you know to

say uh ji this accident I'm really glad that I can say this yeah so this is about the our our first uh wave of

launch. A lot of people talking about us

launch. A lot of people talking about us is super viral. But you know it's not just about viral. We also get a lot of traffic that we that we can't handle.

Yeah. Because you know at the very first uh we just prep we in our original expectation we saw in the first months

we may have in our wildest gas we may have 100k users. So we only prepare very like limited resources

and the the virtual machines very limited resource for the launch but after the first day we got huge traffic.

So we are really busy on like how to scale the whole system and uh we even forget about to put pricing before the

launch. So in the first maybe like 22 21

launch. So in the first maybe like 22 21 or 22 days man is running without pricing which means you know like at that time we only have $8 million in our

bank account and every day we are burning nearly 500k a day which means if we don't have a wheat system we don't we don't we don't have a invite in

invitation code system our company will be broken in just one or two days.

That's why, you know, we keep the weight system for a month is because we're under like huge heavy traffic. We can't

handle and we don't have a pricing. And

uh just uh like 23 days after the launch uh we finally get the pricing system online and then we can uh open the whole system to public because now people can

pay. Yeah. So that's kind of like the

pay. Yeah. So that's kind of like the the the first months. Uh it's it's really crazy months. Yeah. And uh after

launch the uh manas uh we we we decided that because you know man users is distributed all across the world we decided that maybe we should meet them.

So we we we we decided to go out. The

the first country we went is South Korea and uh because you know Anthropica invited invited us there and uh that is

kind of like March 19th which is just 14 days after the launch but Anthropica is already put manas in the worldass coding

category. We just you know we have a

category. We just you know we have a very small startup at that time and we are see we when we're sawing that seeing that as you put us here along with all

these very famous projects we just feel very happy. Yeah. And after South Korea

very happy. Yeah. And after South Korea we went to uh Bay Area and attended the Nvidia's GTC and our chief of staff Parker uh is wearing this t-shirt with

me which is we made this t-shirt. Yeah.

We burn $1 million tokens in the first 14 days. Yeah. Do you want to know why?

14 days. Yeah. Do you want to know why?

And a lot of people, you know, uh coming to us and talk about, okay, what do you guys are building? And uh uh in that GTC

event, uh we uh I think the the the happiest moment is when I'm I'm working uh in different booths and I found

there's one booth from H2O. they they

also making like enterprise agent and they are showing their Gaia benchmark is outperforming us. We at the second place

outperforming us. We at the second place but I'm super happy because that is March 20th which is just 15 days after our initial launch and all the players

are already put manus at their competitor. So I'm I'm very happy. I

competitor. So I'm I'm very happy. I

just want to be at the second place of every company's benchmark. Yeah.

[laughter] Yeah. super happy. Yeah. And after that

Yeah. super happy. Yeah. And after that we made our many user meetups in San Francisco and yeah and he's from E2B which you know supporting the uh

infrastructure of manas is virtual machine nature and we were like telling all all the agent founders in San Francisco about why we built this tool

and uh how do we think about the uh aging architecture. Yeah, a lot of

aging architecture. Yeah, a lot of events and I even went to Los Angeles to UCLA uh and uh hold an event at the boat

uh at the Boer Hall which is the birthplace of internet. The first

message of internet is from USA to Stanford. So we just give a speech at

Stanford. So we just give a speech at this laboratory the voter hall birthplace of the internet. It's really

unforgetting experience for me. And

after the west coast, we go to the east coast, we went to MIT, of course. Why we

went to MIT is because you know if you are looking at this in MIT, yeah, you will see here is men at manus. It's a

old Latin words which means man and hand. Men's is a man and manus is a

hand. Men's is a man and manus is a hand. Why we choose this name manus? is

hand. Why we choose this name manus? is

because we we think all for all the past two years all these frontier model labs they are trying to build the brain but we think if you want to make real impact

into the physical world you don't you don't you can't just build the brain you still have to build the hands for the AI to take actions so we decide to build

the manners which is the hands for LMS yeah so that's why how we choose this name from MIT's model So I'm I must go

to MIT. Yeah. And we we give so many

to MIT. Yeah. And we we give so many speeches at MIT, Harvard and uh this is our first Japan event. Uh this is in

Saudi Arabia and uh the most the happiest thing for me is when I was in Paris attending Vivate. I was invited to a dinner and luckily Tony Fedel is

sitting beside me. Yeah. And you know I used like everything he designed from the classic iPod to nest you know like

everything he designed is is just touch my heart and uh thanks to manness I have the chance to sitting next to him. Yeah,

it's really our owner. And uh after that uh we relocated to we relocated the company to Singapore and we made some

ads at Singapore uh San Francisco and I I guess you may see some of them. Yeah.

And also you know our our our friend company you know like E2B they also put some ad for us in San Francisco. Yeah I

must say thanks. Yeah. So that's kind of like the uh the whole thing you know after the launch is we got a lot of traffic we got a lot of users and we got

a a lot of attentions in the industry across the world like in the past nine months is we did a lot of improvements

for our product like you can see here.

Yeah, like this sites from uh only include the features we released from July to December and uh as you can see there uh in July we have data

visualization new slides since even this slides is generated by manus itself. I

just asked manus to generate a slide showing all these fetches and updates manus has released since dr. So mass just delivered this site for me. So I

can show it to you now. And in August we release is a very important feature. We

call it wide racer. Find all YC21 founders email for me. And as you can see, manus decided to search all these things first and then browse the Y

combinators website clicking some element to filter uh that batch and then uh scroll scroll scroll and extract all

their names and instead of just looking all these companies information one by one, manus used the wide research feature to find all these companies info

and their founders. email in one shot.

You know, like this is something only man can do to do some simple researches and investigations on hundreds or thousands of different subjects like in

this one. It's one it's 336

this one. It's one it's 336 companies information and they are all running concurrently. So which may make

running concurrently. So which may make the results is very fast and it's very reliable and after that man after doing the wide research manusc will give you

the original spreadsheet of the wide research result. Yeah. And also a report

research result. Yeah. And also a report about all these founders information.

Yeah. So this is something we call wide research. And also we have like like

research. And also we have like like another very cool feature. I really like this one which is mail manus because you know you don't have to you don't have to stay inside manus to use manus we have

this feature we call it mail manus as here you can just set up your own manus board email mine is highcloud manboard you don't have to send emails to this

address because you know this can only respond to my own email yeah so here you can see an example yeah sometimes I will

get some invitations to some event and I have to do the uh RSVP you know by clicking that link and fill out some form and after using manus's mail manus

feature I can just forward this email to my manus bot email and say say what the task he should do for me which is register this event for me and the manus

will okay reply to me okay I received your task and now I'm going to the uh registration website to do the registration for you And after the task

is done, manus will reply into your inbox. And in the whole time, you don't

inbox. And in the whole time, you don't have to jump into Manus. You can just stay inside your email inbox and get all things done for you. It's super cool,

very convenient. And in October, we

very convenient. And in October, we release manus 1.5.

uh in that version we have so many features but the most important part is about before 1.5 we can only develop static web pages but after 1.5 we can

develop fully functional websites that's with a backend service with databases where I with native AI capabilities this is just something I think just seen

things before yeah like this one you see I asked man to build a gumpla collection system for me and this is something mis build for Yeah, which is uh I can in

here I can just mark which model I own.

I can just add it. Yeah, this is with the fully functional and uh if you uh are browsing this website, you'll see you can register this website by

yourself. You have your per your

yourself. You have your per your personal profile. So you can track your

personal profile. So you can track your own progress. This is not static. Yeah,

own progress. This is not static. Yeah,

this is something with the back end and the databases. Yeah, it's very powerful

the databases. Yeah, it's very powerful feature and uh uh after the mass 1.5 uh in the latest we just released some

other things too. Yeah, like uh in like maybe two weeks ago we released minus 1.6 and in that version we also have many things like the max mode. Yeah, the

max mode it's very powerful and we have the mobile development. uh we are not just can just build you know just a website we can also build Android app

that you can submit to Google Play Store. This is really wonderful when you

Store. This is really wonderful when you are looking back at the whole year and by deliver the best agent experience to

our users. We're not just getting

our users. We're not just getting traffic uh you know or something else we are also getting real revenues because you know in this AI trend we believe

that people will use our money to vote for the best AI product. I think that's you know why we can achieve the 100

million AR in less than nine months and I know you know a lot of people are talking about fake AR but we we never believe that the fake AR will work so I

think maybe this is the f first time I share the screenshots you know about our real AR yeah this is a screenshot from

our strive at that day you know when we uh when We launch our 100 million at the December 17th. Yeah, this is our lumber

December 17th. Yeah, this is our lumber from Strive. This is our lumber from

from Strive. This is our lumber from revenue cat because we also have some MR from iOS and Android. We use revenue CAD to track this MR. So if you combine

these numbers together and you will see yeah this is our AR number and this year is like really amazing you

know like both for us and for all the users of manas I just want to say what a manus year and we're really like I I we

just reached our first milestone but we're really excited about the next year like what's going to come in 2020. 26

and uh I believe you know next year there's going to be more and more tasks you will leave it to manness and let manus handle them for you. So I just

want to say that again it's like leave it to manness. Thanks for watching.