ByteDance's Trae Work Goes Free for All, Bundling Seed 2.1 Pro and a Full Coding Agent

I heard that ByteDance's Seed 2.1 Pro is awesome again this time!

I've been looking for ways to use it and found that only their Trae Work offers immediate free access. So let's get Work up and running.

These days, Codex seems to have quieted down. Meanwhile, similar domestic software is becoming active. We've previously introduced Zhipu's version of "Codex" — ZCode!

Today, let's experience ByteDance's version of "Codex" and the latest Doubao Seed 2.1 Pro model!

Key point: All models on Work are free for everyone!

Yesterday during the day, 3.1 Pro was still unavailable, and there were over a thousand people in the queue. By evening, it was unlocked and running very smoothly, even handling multiple projects simultaneously.

Let's see how to get started.

Immediately go to the official website, download the installer, double-click to launch, and click Next all the way:

After installation, it looks like this:

Trae Work clearly follows the Codex-type approach, and it has also learned from Claude's structure. It divides its main functions into Office (Work) and Programming (Code) modes, and also supports mobile remote access.

Anyway, it has all the features you'd expect from this kind of general-purpose agent. ByteDance's product experience is undoubtedly top-notch. Judging solely by the software's smoothness, it's definitely among the best domestically!

Below, I'll present this software in three parts!

1. Mobile Remote

Although I might only use it a few times a year, I'm really fascinated by this kind of remote functionality. I gave it a rough try, and it was also very smooth. Especially the seamless synchronization of conversations. Codex requires multi-device scientific internet access, whereas Trae Work can be used directly without any proxy, making it very flexible and smooth.

Once the mobile version is installed, it achieves almost seamless synchronization:

There's a remote configuration entry on the desktop version; clicking it brings up the configuration interface.

Actually, not much configuration is needed; the main thing is to enable control.

This step has two options: one allows mobile devices to control this device, and the second keeps the computer awake. I usually enable both; my PC never shuts down. Nowadays, most laptops are also just closed directly!

Click finish, and then all conversations and progress will sync in real-time!

Its synchronization is very smooth! Message notifications, task lists, and detailed conversation information. All of this syncs seamlessly.

Compared to ZCode, Trae Work is clearly more mass-market oriented with a better experience. ZCode is obviously more professional and more suitable for developers.

My previous remote tools were basically the GPT mobile app and Claude mobile app, which worked quite well. But using them with domestic phones and computers was a bit troublesome. Now I can use my Mate phone and Windows machine, taking the Trae route.

2. Office Work

With the popularization of AI and the enhancement of model capabilities, these agents are no longer limited to programmers. Everyone is shifting from Coding Agents to Work Agents, competing for the office worker market.

So ByteDance is heavily promoting its Work feature! The name above is also Trae Work!

Let's see how this Work feature performs? What can it do?

After opening the software, it defaults to Work mode. The interface is very simple, just an input box.

In the bottom right corner of the input box, you can select the model. The default is auto mode; you can click to select the latest Seed 2.1 Pro.

At the bottom, four common use cases are provided. Let's try them all today!

Because when I first used it, the 2.1 version was still locked, so I directly chose auto mode to create. Let's see the results.

The first example is "Web Page Reading":

Study this paper https://openreview.net/forum?id=WE_vluYUL-X and generate a paper review that is easy to understand, clearly structured, and illustrated with text and images.

Its execution time and results are as follows:

The final deliverable is a webpage, which you can click to view:

This page is quite well done, with both text and images!

A very critical part here is that it can automatically generate appropriate images based on context. Since ByteDance is top-tier domestically in multimodal, image generation, and video generation, this Work feature is very "Work" indeed. Similar to how Codex comes with its own top-tier image generation capability, it directly leads competitors by a level.

The second example is "Research Analysis":

Research the differentiated advantages and development paths of current mainstream short-video platforms, and compile a presentation draft for reporting. The research scope includes platform basics, user scale, content ecosystem, recommendation mechanisms, commercialization models, and representative cases. Focus on comparing differences in user groups, content types, and growth strategies across platforms, and summarize their successful experiences and future trends to provide reference for product or market strategy.

Click the use case and send directly:

This task took 13 minutes. It created a 21-page PPT.

It includes a large number of visual elements:

• ECharts bar chart (China platform MAU comparison)
• ECharts radar chart (five-dimensional user profile comparison)
• ECharts pie chart (commercialization revenue structure)
• Data comparison tables (platform profiles, algorithm mechanisms, overseas comparisons)
• AI-generated professional background images (cover, chapter dividers)

This is ByteDance's strongest domain, and since this is an official demo, the effect is naturally stunning!

The report is presented as a web-based PPT, with clear thinking, comprehensive angles, rich data, and excellent illustrations!

If it can achieve this level in all domains, it would truly be very "Work"!

The third example is "Data Mining":

Mine and analyze the global AI industry's market growth data over the past three years. 1. Including investment scale, number of enterprises, technology directions, and main application areas. 2. Data sources can include industry reports, public statistics, and market research materials, requiring source annotation. 3. By organizing and visualizing the data, analyze the development trends, hot tracks, and potential growth opportunities in the AI market in recent years.

Similarly, click the use case and send directly:

This example took about 7 minutes:

Based on public data, it sorted out the core indicators of the global AI market from 2023 to 2025.

It ultimately formed a web-based report.

The previous part had automatically generated images; this part contains a large number of "interactive tables".

This kind of report with automatically mixed text, images, and charts is quite pleasant to read!

Finally, there's one more example: File Organization!

I won't demonstrate this one because my files are too many; there's order in the chaos, and once you organize them for me, I won't be able to figure them out.

Judging from the demonstrated examples and results, Trae's Work is clearly different from many conventional Work tools. It actually leans towards more professional users.

It doesn't simply focus on making a Word document or a PPT, or handling trivial daily tasks like Doubao does. This Work is clearly knowledge and information-intensive Work!

Because the entire report is very long and contains a lot of content, I didn't show everything. But from a rough preview, from my not-so-work-oriented perspective, it's still quite professional.

Moreover, I was using the auto-selected model at the time, which was likely a previous generation version. This morning, I found that 2.1 Pro in Work has also been unlocked, so the new version will certainly be better.

Actually, the interface of these agent software is very simple; the key is still its model and how to control the model to perform better!

This can only be known after in-depth use.

Especially, how it performs in scenarios beyond the official examples is still a big question mark ❓

Today is just an experience, not a review, so I won't elaborate.

3. Code

The Work feature is finally finished. Next is our most familiar Code feature.

Generally, small companies might need to choose between Work and Code, but as a big company like ByteDance, of course, they want both.

Let's first look at the main interface of Code:

The interface is similarly very simple, just an input box, a model selector, and some use cases at the bottom. These use cases mainly include app development, project understanding, game ideas, and utility scripts. I actually ran all of these examples.

Yesterday during the day, 2.1 Pro in the Code section wasn't unlocked yet. I ran three examples using 2.1 Turbo.

In the evening, I found that Pro was available, so I re-ran all of them!

I won't talk about the results first; I'll just say that the token allowance is really very generous.

At 11 PM, running several test projects simultaneously, I actually didn't encounter any rate limits.

During the day, I complained that ByteDance was too stingy, but at night it was different. I ran a bunch of projects, no idea how many tokens were burned, and it felt great.

I remember ZCode gave away 5 million tokens at the time, which might be used up in one project.

If it were Kimi, even with its paid plan, running these examples would probably take several days. One example would consume several of its 5-hour cycles. With so many examples, it would basically consume a week's quota.

Alright, enough chatter, let's see the results. Let's look at its first example, "App Development":

The requirement for this example is as follows:

Develop an online education platform supporting multilingual learning, covering mainstream languages like English, Japanese, and Korean.

The platform needs to create an immersive language learning experience, providing the following capabilities:

1. Tiered curriculum system

2. Interactive learning modules (vocabulary memorization, grammar exercises, speaking practice, listening training)

3. Learning progress tracking

4. User registration and login support

5. Personalized learning path recommendations

6. Community interaction and achievement incentive system

This example consumed a total of 1 hour and 8 minutes.

The final product is as follows:

I didn't expect Doubao to be able to work continuously for so long; it's really impressive progress.

This is a project based on React 18 + TypeScript + Vite + TailwindCSS, with the following directory structure:

src/
├── components/layout/     # Layout components (Navbar, Footer, MainLayout)
├── pages/                 # 13 page components
│   ├── Home.tsx           # Homepage
│   ├── Login.tsx          # Login
│   ├── Register.tsx       # Register
│   ├── Courses.tsx        # Course list
│   ├── CourseDetail.tsx   # Course details
│   ├── Learn.tsx          # Learning center
│   ├── Vocabulary.tsx     # Vocabulary memorization
│   ├── Grammar.tsx        # Grammar exercises
│   ├── Speaking.tsx       # Speaking practice
│   ├── Listening.tsx      # Listening training
│   ├── Dashboard.tsx      # Learning data
│   ├── Community.tsx      # Community interaction
│   └── Profile.tsx        # Personal center
├── store/                 # Zustand state management
├── data/                  # Mock data
├── types/                 # TypeScript types
└── utils/                 # Utility functions

This structure is very clear, and this tech stack is currently what AI is best at. Anyone coding with AI has probably seen this many times.

We can see that after development, it also performed "quality assurance":

• ✅ TypeScript type check passed
• ✅ Production build successful
• ✅ Dev server running normally
• ✅ All page routes accessible
• ✅ Responsive layout adapted

Then it generated two documents: one technical architecture document and one PRD requirements document.

The technical document includes chapters on architecture design, feature descriptions, route definitions, data models, core functions, module descriptions, performance optimization, etc.

The PRD document includes chapters on product overview, core functions, core processes, user interface design, etc.

After all, it's an official demo project, so it's done very comprehensively.

The upper right corner of the software also displays a complete to-do list.

The lower right corner shows context, skills, files, and other information!

Let's look at the results. No amount of talk beats seeing the results.

Homepage:

Learning function page:

Registration and login functions:

Footer effect:

Specific course page:

After all, it wrote for over an hour, so this example is very complete.

First, there are no basic code errors!

Second, there's no layout confusion, and the design is decent.

Then, the functionality is quite comprehensive. The content of each section, including registration/login, header/footer, all visible functions basically work.

Judging from this example alone, the coding capability is quite solid! Of course, it's an official demo, everyone understands that. I won't elaborate for now!

Then I experienced its second example, "Project Understanding":

Coincidentally, I had just developed a project, so I let it understand this project and generate a Code Wiki!

It took about three minutes to finish writing the Wiki:

There's a lot of content, so I'll just take a screenshot to give an idea:

This is good!

It has a good chapter division, and each chapter is written according to a relatively good standard.

This part would be very tedious for humans to do.

Over a decade ago, I developed a relatively large project. I wrote hundreds of pages of delivery documents at the time, spending several weeks, almost going crazy. It was much more painful than writing code. After finishing, I didn't even have the energy to review it once, and it was soon printed into a thick book~~ I don't know how many typos there were. Fortunately, it has nothing to do with me now!

In the future, this kind of drudgery can be handed over to AI! The document quality of Trae Code is quite good; it should have undergone specialized optimization training! If this feature can handle documentation chores, that's not bad.

Besides the above two examples, I also experienced the third example, "Creative Game":

I won't comment much on this one. I'll just say, "If you're bad, practice more"!

That's about it for the software-related content!

I deliberately avoided the model part in this article. I just approached it from an application perspective, focusing on showcasing the Trae Work software.

This is also my first time using this software. Today is just a first impression. First encounters are generally better; once you get familiar, only complaints remain.

I've already used many similar software.

Overall, Trae Work is very easy to get started with. It's just an input box, select a folder, choose a model, enter your requirement, and hit Enter.

The Trae series actually started very early, the earliest among all domestic similar software. You might have seen this icon (the green one) in my screenshots before.

I've always been a Trae international version user. Initially, I went there to freeload. After they started charging, I topped up $100 to get Claude 3.7, boarding the "pirate ship," and then Claude 3.7 was immediately taken down! Hahaha~~ So don't casually buy annual passes for AI products.

They initially made a VS Code-like IDE, then went all the way to Solo, and now to Work. I experienced the first two immediately, but this Work is my first time using it.

You can always trust ByteDance's user experience; after all, it's the tech company with the strongest product capability currently.

Moreover, this Trae walks on two legs, both domestic and international, very comprehensive.

The most magical point is that you can directly use the GPT and Gemini models from the international version domestically without scientific internet tools. It's a pity that Claude has been cut off; otherwise, it would be one of my main tools!

Because I bought an annual membership, I've been using the GPT inside for chores, like checking file contents.

Next, I might also abandon the IDE version and switch directly to this Work version.

As a purely domestic user, Trae Work seems like a good choice too!

After all, it has everything you need: Work features, Code features, and mobile remote functionality.

And it includes almost all domestic models, besides its own Seed series, there are also the latest GLM-5.2, Kimi-K2.7, MiniMax M3!

Note, usually these models are all free (queuing for popular models is inevitable).

I think they won't be able to hold out much longer and will gradually start charging.

That's it. I've already run many coding test projects on it. You might see the results quickly this afternoon, or tomorrow at the latest!

Let me test how much "Sis Dou" really weighs? Is the ARENA leaderboard reliable? Can it really wrestle with Opus 4.7?

I'm generally more lenient with software, but stricter with models!