Screw Agents, Workflows, Context, Skills, and Connectors

TL;DR: Saying what you want and getting it done takes a computer built for AI, not for you: Pine Computer, in the cloud.

Imagine someone from fifty years ago stepping out of a time machine today. They’d be stunned that most of humanity spends its working hours staring at glowing screens. In 1976, there were about 40,000 computers in the entire world. Apple was founded that same year.

Now imagine someone from fifty years in the future stepping into today. They’d be just as stunned, for the opposite reason: why are people still sitting at their own computers, doing the work themselves? In their world, AI does the work. If you use Codex or Claude Code every day, you already feel it.

The question isn’t whether. It’s how fast. And one problem stands in the way.

The 0.7% problem

AI needs a computer. Ideally your computer: your context, your tools, your accounts. That’s why OpenAI, Anthropic, Meta and xAI are all racing to ship consumer desktop apps.

But most people don’t want AI installed on their computer.

That sounds wrong if you live in San Francisco, where Codex and Claude Code feel universal. Step outside the bubble. Across the US, their combined active users add up to roughly 0.7% of the population. Instagram and Facebook reach around 60%.

And these aren’t programmer tools anymore. They’re effectively what “AI” means today, capable of general work. The best AI tools in the world, backed by companies with practically unlimited resources, in the world’s most developed market: 0.7%.

Pick your poison:

AI belongs in the cloud. AI also belongs to you.

We believe the work of the future will run in the cloud.

The cloud era already proved it. Almost no product runs its own on-premise servers anymore. Virtualization made the cloud the obvious home for backend work.

But AI isn’t backend work. AI is personal. It touches your data, your software, your permissions, your logins, your accounts.

So we need both: the reach of the cloud and the intimacy of a personal computer.

That’s what we’re launching today: a computer built for AI, in the cloud. We call it Pine Computer.

A new species

Kubernetes is mature. Why can’t AI just run in the cloud like any other workload?

Because an AI computer has to be all ten of these at once:

1. Dedicated. It belongs to one person. It can’t be shared like a backend service, because it holds your data, your privacy, even your login credentials and assets.
2. Secure. As secure as your local computer, or more.
3. Identity-aware. Your local computer holds your logins; that’s why work gets done there. A cloud computer must work just as easily, with passkeys and existing login sessions.
4. Permissioned. When no approval is needed, the AI just works. When it is, such as entering a password or signing a final authorization, a human steps in.
5. Parallel. Spin up many computers at once, each sized to the job.
6. Disposable. Created on demand. Shut down, or deleted, the moment the task is done.
7. Persistent. What belongs to you, rather than to a single task, outlives any one computer.
8. Fast. Boots in seconds. Runs no slower than a personal computer; in fact, no slower than a person at the keyboard.
9. Equipped. Purpose-built environments for each kind of work. Audio and video production gets ffmpeg, editing software, and access to text-to-image and text-to-video APIs (the keys never live on the machine). Office work gets a full office suite. Every computer gets the essentials: small local models, PDF tools, Python. And all of it is controlled: anything extra installed, any port opened, happens under a strict security policy.
10. Watchable. A live streaming screen, so people can see progress in real time.

Nothing that exists today checks all ten. This is a new species.

Where the plumbing goes

Today, getting AI to do real work means assembling the plumbing yourself. You build an agent. You wire up a workflow. You stuff context into prompts, install skills, and set up connectors to every service it needs. None of that is a law of nature. It’s what you’re left doing when AI has no computer of its own.

Give it one, and each piece becomes part of the machine:

• Context persists. Your sign-ins, and everything else that belongs to you rather than to one task, carry over from one task to the next. Nobody re-explains the world each time.
• Skills come equipped. Document tools, media tools and the essentials ship with the computer. And it learns new skills as it works, from its own runs or by watching a person, and saves them on the computer.
• Connectors become logins. Most websites have no connector and no API. Using the site itself, signed in as you, the way your laptop does, is the more general way. Connectors were a workaround for AI without a computer of its own.
• Agents and workflows collapse into one sentence. You don’t assemble a loop; you hand the computer a goal.

What happens when you stop building for humans

Here’s what we didn’t expect. Once a computer is built for AI, primarily if not exclusively, we barely have to care about the human user experience. That freedom turned out to be a superpower:

1. Faster. We built a runtime and harness that reach deep into the operating system and the browser, binding to event notifications and watching for change. The ReAct loop runs faster, and every step gets smaller, more precise context.
2. More than an order of magnitude cheaper. On a computer and runtime purpose-built for AI, a lower-cost model can carry work that would otherwise take a frontier model on a PC. On SaaS-Bench v1.1, Pine Computer with GPT-5.6 Luna spent about one twenty-sixth of the model-token cost of Opus 5 with Claude Code.¹
3. Your own computer stays clean. Environment setup, downloads, intermediate artifacts: all of it lives in the cloud. Let it go. Keep only what matters.

A tale of two clips

Yesterday I wanted to catch up on a TV interview I’d missed: a transcript, a short briefing, and a punchy 30-second vertical clip.

Claude Code, with Opus 5, did it on my laptop. It set up a Python environment and downloaded a 1.6 GB Whisper model to transcribe locally. It asked for my approval at every step. If I weren’t an engineer, I’d have had no idea what it was doing, and I would never have dared to let it. Eighteen minutes later I had my clip, and my laptop was 1.7 GB heavier. Even with Claude Code’s consumer subsidy, it burned a good chunk of my $100 Max plan credits; at API list prices, the model alone came to $5.34.

Then I sent the same prompt to Pine Computer through the API, with GPT-5.6 Luna. Pine Computer already had ffmpeg, a transcription API, everything it needed. Five minutes and 24 seconds later, the transcript, the briefing and the clip were on my laptop, 14 MB in all, and one more API call shut the computer down. The model cost $0.066.²

Nearly 500 GB of my 1 TB MacBook is probably this kind of debris: intermediate files, videos, Python and C++ libraries, software I used once and never again.

Multiply that by everyone who isn’t an engineer, and you’ve found the 0.7%.

One line

Look how far we’ve come.

Then:

response = openai.Completion.create(
    engine="davinci",
    prompt="Write a tagline for an ice cream shop:",
    max_tokens=20
)

Now:

response = client.responses.create(
    model="gpt-6-astra",
    instructions="Solve the problem and verify it with code.",
    input="Analyze this dataset.",
    tools=[
        {
            "type": "code_interpreter",
            "container": {"type": "auto"}
        }
    ]
)

With Pine Computer (its SDKs are Ruby and Go today):

computer = pine.create_computer
session  = computer.create_session(name: "main", browser: true)

session.agent.run("Use this video as a reference and make me a similar one: https://…")
session.agent.events { |event| break if event.terminal? }

File.binwrite("results.zip", session.download_artifacts_zip)
computer.stop

The rest is setup you write once: a computer, a session, the results, and a stop. The work itself is one line, one sentence of intent. The computer does the rest.

For consumers, for app users, for anyone who just wants things done: screw agents, workflows, context, skills and connectors. You’ll never have to think about any of them again.

What’s next

This is a paradigm shift, and a shift this big won’t be built by one team or one company. So we’re doing it in the open:

• We’ll publish our design, implementation and protocols, piece by piece.
• After an audit by our security team, we’ll open-source the harness and runtime.
• We’ll open-source how we handle logins and credentials, so you don’t have to take our word for it.
• OfficeVal results are next, and we plan to publish our evaluation tools and a playground.

We need the open-source community, and AI and infrastructure companies across the industry, building this with us.

We started Pine Computer because we believe every AI deserves a computer of its own. Come build it with us: join the waitlist.

Questions? We answer the ones you’re already typing in Introducing Pine Computer.