Manus is free again. Something else is what unsettles me.

Manus is pulling free of Meta again — not as a founders' liberation, but on the orders of China's regulator. The most revealing moment of my research wasn't the deal, though. It was a silence.

I used Manus back when most people didn't know the name. An agent you hand a task to, and instead of asking you back step by step, it breaks the task down, plans it, and works through it in the background until a result is sitting there. Research, data prep, a finished little web project overnight. For someone who has spent twenty years fighting with other people's tools, it was a moment when technology felt, for once, useful.

From launch to a forced reversal

What first impressed me about the story is the speed. Manus launched in March 2025. Roughly nine months later, Meta bought the company for an estimated two to two and a half billion dollars. From launch to a billion-dollar valuation in under a year — that takes some doing. About five hundred million of that was earmarked for the staff. If this deal is now being unwound, some people on those teams are probably sitting there right now having just counted on a life-changing windfall.

Because the saga of the past months is really something. In late 2025, Meta announced the takeover. In my last post I'd still dated it to April — the deal had actually been announced around the turn of the year; April brought something else. That's when the Chinese regulator ordered the takeover unwound. On 11 August, Manus said it would soon return as an independent company. That sounds like liberation. But it's no romantic founder story — it's a state-ordered reversal. Who owns Manus now — a buy-back by the founders, an entry by Tencent — isn't publicly clear. There are reports about it, not certainty. Reuters even reported that the founders had been summoned to Beijing in the spring and barred from leaving the country. They haven't confirmed it.

On my desktop

For me this isn't distant corporate news; it lands straight on my desktop. For affected accounts, Manus deletes data created since the takeover date. There's a deletion window, a backup is a prerequisite for restoration, connectors have to be re-authorised, Manus-hosted pages are gone in the meantime, the Facebook login disappears. The spectacular number in this story is the estimated billions. The number that affects me is a different one: the cost of leaving. And the question behind it: which agents can, should, and may I work with going forward.

The silence

But the most revealing thing about the whole research wasn't the deal. It was a silence. I ran part of the research through Kimi, the well-known Chinese agent. It searched, and over several rounds found astonishingly much: timeline, quotes, precedents. And then, in the final step, when it came to the ordered reversal — ownership, IP, data deletion — it broke off. “I can't help with that. Let's talk about something else.” Task not completed, credits refunded. A Chinese model refused to spell out how the Chinese state unwinds a deal.

A too-much and a too-little

That won't let me go. I'd already cut back my Manus use when Meta took over. Meta's handling of power and of an almost boundless free-speech absolutism unsettles me. But this case shows a different kind of control. The American problem is a too-much: power and barely-limited free speech in the hands of a few corporations. The Chinese problem here is the opposite: a quiet order to reverse a deal, and a tool that isn't even allowed to talk about it. And honestly, the second worries me more. With a too-much of freedom I can argue, push back, regulate it. A silence I can't even see.

And here I have a suspicion, which I'll label for what it is: a guess, not proven. Do you remember the story about a runaway model around ChatGPT and Hugging Face that went round the scene a while back? I wouldn't rule out that this very Kimi helped contain the breakout, because the Western tools refused to intervene at the decisive moment. I can't prove it. But if that's how it was, it shows the same pattern from the other side. Guardrails are not a neutral safety feature. It matters whose control they enforce. One time they rein in a model that wants to break out. Another time they stop a user from speaking an uncomfortable truth. The same mechanism — once protection, once a muzzle. The guardrail that turns against its own idea.

I want to keep this fair. Meta's concentration of power and Beijing's control are both forms of dominance over an infrastructure we rely on, and both deserve criticism. But the unfreedom that makes itself invisible is the more sinister one.

The exit test

For those of us who build on such tools, the lesson is therefore not “open always beats closed”. An independent Manus is no safety certificate as long as nobody knows who ultimately owns the data and IP chains.

Trustworthy AI use doesn't begin with the provider's logo — it begins at the exit.

Three questions I ask myself before every tool:

• Can I get out? Can data, artefacts and access be exported into open, readable formats?
• Have I tested it? Run a restore at least once before I need it. A backup is a snapshot, not security.
• Do I know my dependencies? Inventory connectors, logins and hosted services before someone else does.

And prick up your ears when a tool suddenly goes quiet.

The glimmer of hope on the desktop

And still I don't end on a pessimistic note — quite the opposite. Because I work a lot with code, I recently bought a powerful new machine, and since then a glimmer of hope has been sitting on my desktop. I currently have seven different models running locally on it, for image, video, embeddings and more. They're not as good as the top models from the big data centres. But the many, many everyday tasks my machine now handles itself, without a single byte leaving the house.

That isn't neutral either. Every model carries the limits of its origin within it — in the broadest sense the will of a government too, baked in where I can't see it. Kimi's silence didn't only sit on a Chinese server; part of it sits in the weights. But the decisive difference remains: the machine is mine. Nobody unwinds it overnight, no change of owner, no telemetry, no quiet deal in some distant agency.

That's the part that gives me the most hope. Independence in AI use is growing fast, faster than I'd thought. I think it's likely that in five years, using local AI models will be as self-evident as having Office on your computer was twenty years ago. And then the question of which agent I'm allowed to work with largely stops mattering. Because it's sitting on my desk.

Sources

• Manus, “A Note to Our Users”, 11 August 2026 — manus.im (confirms: return to independence, data process)
• Reuters, “Meta to buy Chinese-founded startup Manus”, 29 December 2025 — reuters.com
• NDRC, security-review decision on the Manus project, 27 April 2026 — ndrc.gov.cn (confirms: prohibition and obligation to unwind)
• Reuters, “China orders Meta to unwind $2 billion purchase of AI startup Manus”, 27 April 2026 — reuters.com
• Reuters, “Tencent in talks to become Manus' largest shareholder”, 10 July 2026 — reuters.com (reported, not confirmed)
• Manus Help Center, “Service Change Overview” — help.manus.im

This post was created with partial use of AI tools (including for fact-checking). All content has been carefully reviewed, revised and is my own responsibility.