An AI Agent Landscape: Products in China, the US, and the Open-Source Community ​

We · October 2026

This is the second article in the "AI Agents & Local Automation" series. The first one explained, in plain language, what separates an agent from a chatbot. This one is about what is actually out there: products from the major Chinese companies, products from the major American ones, where Europe and Japan actually sit, and the open-source project that defined the whole category. The names get dizzying fast. What we want to do is put them on one map.

AI moves fast. The tools, capabilities and numbers in this article describe where things stand right now, and are for reference only. One thing up front: this is a map, not a leaderboard. We are not ranking anything and not calling anything the best. We do two things—separate the categories, then give you a few questions to answer.

Start with the two axes ​

Looking at these tools, you only need to ask two questions.

Question one: where does the work happen?

• On your own computer. It reads and writes the folders you authorize, clicks around in your browser, runs commands on your machine.
• On the vendor's server. It works inside a cloud virtual machine and keeps going after you close the laptop.
• A mix. Interaction happens on the desktop client; scheduled and long-running jobs live in the cloud.

Question two: where does the model come from?

• Only the ones the product ships with. Want a different model? You follow their price list.
• Bring your own API key—BYOK in industry language. You use model credits you bought somewhere else.
• A model running on your own machine. Connect it to a locally deployed model, so neither the data nor the compute leaves the computer.

These two axes matter for one reason each: the first decides who ever sees your files, the second decides whether you can take what you have built with you if you switch providers in three years.

One thing that gets confused a lot: different modes of the same product land in different cells. An office agent may run its desktop mode on your machine and its cloud-computer mode on a server. Look at the mode, not the product name.

First, a sense of scale ​

Before learning any names, here are a few numbers with their sources attached. They answer one question: is what sits on this map a hobbyist toy, or already part of mainstream work?

• The research firm Analysys reported in October 2026 that in September alone, the 20 desktop office-agent products in China together passed 300 million monthly visits—roughly five times the June figure. The busiest were Tencent WorkBuddy at 85.34 million, Alibaba's QwenWork at 46.55 million, and ByteDance's Doubao Work at 30.39 million. Be clear about the unit: these are visits, not users, and the two cannot be used interchangeably. Apart from the occasional figure inside a company's earnings report, this category simply has no authoritative active-user data. That itself is a signal.
• IDC's read on the enterprise side: China's AI agent market was roughly 21.2 billion yuan in 2025, is expected to reach 44.9 billion in 2026, and could exceed 332 billion yuan by 2029.
• Morgan Stanley measured average Chinese large-model API pricing in Q2 2026 at about 4.9 yuan per million input tokens and 21.9 yuan per million output tokens. Those two numbers only make sense read together: a single model call has become so cheap it rounds to nothing, but one agent task calls the model dozens or hundreds of times. Which is why every pricing model on this map has changed shape. Articles three and four of this series follow that thread to the end.

Office and coding agents in China ​

One shared trait across this camp: almost every vendor sells "it runs on your machine, your data never leaves" as the core differentiator.

Three things we noticed:

One: "data never leaves" went from a principle to a selling point. Privacy and data boundaries used to be what local-deployment tools argued for. Now it is the differentiator commercial products use against cloud rivals—config files stored only locally, credentials that never go up, memory kept on the machine. You can find this language in several vendors' own documentation. That is a directional change.

Two: the category is still shuffling hard. Within one year, three products merged into one, two teams were folded into a different entry point, and one vendor's desktop tool published a shutdown notice. Names, entry points and team ownership all keep moving. This is not one company's problem; it is what an immature category looks like. It also means whatever you pick today, you will likely re-decide next year.

Three: pricing has moved to "credits", and cross-shopping is nearly impossible. Most products no longer price by request or by token directly. They use credits, resource points, energy—intermediate units that refill on rolling windows and do not carry over between seats. Tiers reported by media contradict the official pages. If you want to know your real cost, you have to run your own usage for a while. We take that up in the fourth article of this series.

Products from the US ​

Fewer companies than the Chinese camp, but each one is more tangled—usually four or five shapes at once: for coding, for office, running in cloud, running locally.

Four observations:

One: the pitch moved from "how strong is the model" to "how good is the harness." The same model inside different execution shells feels completely different—whether it remembers progress, retries after a failure, and shows you safely which permissions it took. Those are harness properties. It is why everyone is shipping agent clients instead of only shipping models.

Two: pricing is drifting toward "no ceiling." From per-request, to per-credit, and now some vendors bill hosted agents by session-hour. Top consumer tiers commonly sit in the $100–200 per month range, and enterprise moved to "seat fee plus usage at API prices." Flat monthly fee for unlimited use is disappearing from this category.

Three: desktop execution versus cloud execution is not a maturity gap. Microsoft is a good illustration: the enterprise product lives in the cloud (those documents are already in the cloud), the personal task runner uses your own machine. The real split is a trade-off between data boundary and long-task continuity. Which one you want depends on what you are doing.

Four: every company in that table is American. That is not a result of how we picked. It says something about the shape of the category right now—see the next section for why.

Europe, Japan, Korea: the rest of the map ​

One name that is easy to miss. Mistral, based in Paris, took its chat assistant Le Chat, renamed it and turned it into a unified agent platform called Vibe at the end of May 2026. It comes in three modes: an office mode that connects to email, calendar, company documents and collaboration tools to carry out multi-step tasks; a coding mode available as a command-line tool, an editor extension and a web interface, able to write code, test it and submit it; and the plain chat mode it started with. There is a free tier, and paid plans start around fourteen euros a month.

It deserves its own paragraph because its selling point is different from both the US and China. American vendors sell capability. Chinese vendors sell "it runs on your machine." Mistral sells data sovereignty and European-local compute—announced on the same stage were a new inference data center outside Paris, industrial AI partnerships with Airbus and BMW, and citizen-facing government AI built with France, Luxembourg, Greece and Slovakia. In Europe, compliance and sovereignty are not marketing copy; they are features that turn into purchase orders.

The gap should be stated honestly too. The company's own 2026 revenue target is 1 billion euros, while the leading US vendors are operating at roughly twenty times that scale in annualized revenue. Europe has products; what it lacks is a distribution channel that puts them on hundreds of millions of phones.

Now the rest of Asia, where there is an unusually direct set of official numbers. Japan's Ministry of Internal Affairs Communications White Paper, published in July 2026, put generative AI usage at 58.8% in Japan against 93.6% in China, 75.6% in the US and 75.6% in Germany; for "use it almost every day" Japan was 29.9% against 47.9% in the US. More telling is what Japanese users actually run: ChatGPT 36.2%, Gemini 25.0%, Microsoft Copilot 13.3%, Claude 4.3% — local products barely appear on the list.

Korea has its own models and respectable product interfaces—Naver and Kakao both get positive marks from investors. India has enough people to support a large home-grown company, but so many languages that a general-purpose product feels weak in most of them, which is why few local variants have shown up. The common pattern: budgets, models and interfaces exist; the desktop-agent layer does not yet. Japan's AI-related budget for fiscal 2026 is 3.7 times the previous year and the government is running a prize for home-grown foundation models. The money has arrived; the products are still in progress.

We put this section in for one reason: this is not a story about which countries "have AI." The agent category currently requires three things at once—a frontier model, enough capital, and a distribution channel that already sits on hundreds of millions of phones. Companies holding all three exist today only in the US and in China. DeepMind does its research in London but ships products under an American company's name; n8n is developed in Berlin but belongs to workflow automation and was never on the desktop-agent table at all. That is structural position, not a quality ranking.

The open-source community: one project defined the category ​

Those "one-click local install" products in the Chinese camp mostly point at the same open-source project: OpenClaw.

It was open-sourced at the end of November 2025 under the MIT license. What it did, plainly put, was demonstrate a full answer to "what should an agent harness look like": chat apps as the entry point, scheduled jobs and self-waking, file, browser and shell access, capabilities distributed as skill packages, MCP support, and the ability to connect either a cloud model or a local one. Its growth rate was one of the fastest seen in open source. The author later joined a large tech company and the project moved to an independent foundation.

There is a whole field of derivatives and peers around it:

Three cautions in this section:

Open source is not the same as safe. A program that can read, write and execute has, by definition, been given permissions on your machine. Open-source projects do not have fewer security incidents than commercial products. Article three of this series is about risk.

Open source is not the same as free. The harness contains no intelligence; capability comes from whatever model you connect. A cloud model means paying per token; a local model means paying with your own GPU.

Licenses are the easiest trap. n8n uses a fair-code license that is not open source in the usual sense. Dify, Open WebUI, FastGPT, MaxKB and Cherry Studio are all "permissive license plus extra terms"—no multi-tenant SaaS, no removing branding. Personal use is usually fine; anything customer-facing means reading the clauses. Always defer to the LICENSE file in the repository.

One practical note: several categories above are already available as one-click installs on this site—OpenClaw with a graphical front end, a desktop assistant built on the Claude Agent SDK, an AgentScope-based local agent, a chat-bot framework, n8n, MaxKB. For individual users the barrier with tools like these was rarely capability; it was setting up the environment.

Put the launches on one timeline ​

The tables above look like a settled landscape. They are actually events crammed into eighteen months. We laid out every node we could pin down to a month, because that line says more about how new this is—and how quickly it will change shape—than any phrase like "the category is hot."

Three things deserve a pause. First, the starting point is very recent. The standard itself is two years old, and the desktop-agent wave only arrived after January 2026. Second, renaming and merging are normal. One open-source project changed its name twice in a month; three products became one; two teams were folded into a different entry point; one tool published a shutdown notice. Third, both technical standards were handed to the same neutral foundation within ten months. That is unusual in the history of personal computing and mobile, and it means the players now accept that the interface should not belong to any single company.

The standards have already converged ​

This is the section we most want you to remember, because it will outlast every product name in this article.

Two things are now jointly accepted, and they govern different matters:

• MCP handles connection. Wiring tools and data to a model; the official analogy is USB-C for AI. Anthropic published and open-sourced it in November 2024, OpenAI followed in March 2025, then Google, Microsoft, Cursor, AWS and Cloudflare. At the moment of donation to the Agentic AI Foundation under the Linux Foundation, the official figures were roughly 97 million SDK downloads per month and more than 10,000 active public servers. The fifth specification, released in July 2026, was the largest revision ever: the core moved to stateless, with new extensions for long tasks and interactive interfaces.
• Skill packages handle method. A written instruction file, optionally with scripts, that teaches an agent how to do a particular kind of task. Anthropic introduced it in October 2025 as an open standard and donated it to the same foundation in October 2026. Separately, OpenAI, Microsoft, AWS, GitHub, Cursor and Vercel are pushing a packaging format that distributes skills and servers together.

These do not compete: one is an interface, the other is a method. The more practical signal comes from Chinese vendors. Tencent's office agent supports both MCP and command-line protocols in its connectors, and its skill marketplace grew from 78,000 entries in July to over 100,000 in September, with 17 million monthly downloads and 60 million cumulative—while staying compatible with the open-source community's skill file format, so a file can be copied straight into a local directory and just work. Zhipu's one-click install added full MCP support in version 2.0, and ByteDance's no-code builder can connect to other vendors' command-line agents in the opposite direction.

The implication is worth stating plainly: interfaces and methods have crossed camp boundaries, and that is the only long-lived asset an ordinary user can hold on to. A skill file you configured in one product keeps working when you copy it into another. When a vendor shuts a service, what you accumulated travels with you. That is the basis for the closing sentence of this article.

Three questions to place yourself ​

Ignore products, answer questions. We would suggest settling these three before you open any tool list.

Question one: can this work leave my computer?

Client records, unpublished material, personal finances or medical information—lean toward local execution or self-hosting. Want "close the lid, get the result tomorrow morning"—that is a cloud form. This question picks your column on the map.

Question two: do you want advice, or a deliverable?

If a written suggestion is enough, a chatbot is the cheapest and simplest. If it has to land as files, spreadsheets, slides or code, look at what tools the product exposes and what permissions it asks for. If it has to move across several applications repeatedly, look at how well it plans tasks and recovers from failures.

Question three: can it connect to my own tools and models?

Three concrete checks: does it support MCP, does it have a skill-package mechanism, can you bring your own API key or connect a local model. These decide long-term cost and long-term autonomy. A product locked to its own models with no standard interfaces gets harder to leave the longer you use it.

Put the three questions together and the answer is fairly mechanical:

And three audience-specific pointers: if you mostly create content, spend your attention on the office-agent table in the China section; if you study computer science or enjoy building things yourself, start with the lowest-barrier CLI tools and editor plugins in the open-source section; if you are setting up a knowledge base or automation for a team or a household, look at the platform category in the open-source section and pay attention to licenses.

One last thing ​

The shelf life of this map is measured in months. In a single year we saw three products merged into one, a desktop tool announce its shutdown, and two technical standards handed to a neutral foundation. Change is currently the most reliable characteristic of this category.

So here is what we want to leave you with: when you choose, give an extra vote to "compatible with open standards" and "clean license." Products disappear; formats don't. When you cancel a subscription, the skill files you wrote, the workflow you configured, the experience you accumulated running models locally—all of that still works. Tools come and go. What you have built on your own machine stays.