Everyone Gets A Software Company — Benjamin Guo, Zo Computer

Benjamin Guo opens with the product already doing the thing he wants to explain. While waiting to present, he used his Zo Computer to build and publish a site covering the session’s speakers, complete with research, profiles, blogs, and useful links. The observable result is a live page behind the QR code on his slide—not a mockup or a project waiting for a deployment pipeline. This quick build establishes the talk’s working definition of a computer: a place where an agent can research, create files, and publish the result from the same environment. 0:12

That environment reflects Guo’s broader aim for human-machine interaction. He uses Susan Kare’s Finder icon—the human face and machine face meeting in “happy harmony”—as a picture of what good AI tools should feel like. The contrast is today’s ordinary experience: people move among applications, websites, devices, and subscriptions, while vendors raise prices for bundled AI features users may not have requested. The machine no longer feels like one coherent place that belongs to its user. 1:44

Guo calls the underlying structure technofeudalism. The user pays a SaaS provider; the SaaS company rents infrastructure from a cloud provider; the cloud provider depends on chip suppliers. Money and control move upward through that chain, while the user’s information remains split across services at the bottom. Each service controls its own product decisions, pricing, data model, and export paths, so assembling one personal workflow requires stitching together systems with different incentives. 3:17

The ownership problem is more important here than the medieval metaphor. If notes, scheduling, bookkeeping, websites, and customer records live in separate vendor silos, the user cannot easily connect them or replace the interface around them. A product manager at each vendor optimizes that vendor’s business; no one service is responsible for making the user’s whole system coherent. Zo’s proposed answer is a “real home on the internet” where those pieces can share storage, computation, and an agent. 3:47

One early user already runs multiple websites, invoicing, accounts, bookkeeping, scheduling, and notes in this shared environment. The important change is consolidation: instead of copying information among specialized products, the same personal system can access the relevant business files and operate across them. Guo describes the resulting experience as feeling clear, calm, and “fun”—a deliberately emotional standard for infrastructure that usually asks users to tolerate administration. 4:31

A personal cloud is an individually controlled place in the cloud containing the user’s files, AI, and hosted services. In Zo’s implementation, the core is a personal server with AI built in. Because it is a server rather than only a chat interface, it can expose websites and APIs, store files, and run software continuously. The agent and the things it creates occupy the same environment. 5:02

The cloud server can also connect to local devices. Guo’s own setup combines the Zo server and agent with hosted websites, APIs, another personal agent, a laptop, a Mac Mini, and phone access. “Personal cloud” therefore names the whole topology, not one device: the cloud machine supplies an always-available home, while local machines remain interfaces and execution endpoints around it. 5:29

The development model intentionally resembles early web publishing, when changing a site could mean transferring files directly to the server. The agent hides deployment as a separate concern: a regular user asks for a change, and the same system edits the files and serves the result. That simplicity carries a real tradeoff. Direct production changes shorten the path from intent to output, but the talk does not describe staging, rollback, testing, or safeguards for consequential updates. 6:04

The intended alternative is a single environment for personal data, AI, and hosted software rather than constant navigation among separate assistants, devices, and SaaS products. This does not mean no external services are ever involved; later integrations and model APIs remain part of the workspace. The consolidation happens at the user’s operating layer: one server becomes the place where files, tools, and agent actions meet. 6:36

A second early user shows what changes when the personal server becomes a small business’s software company. She replaced services including Squarespace and Calendly with Zo, hosts retreat and personal websites on a custom domain, and keeps the underlying business information together. Her workspace also reflects her own visual interests rather than conforming to a generic SaaS dashboard. 7:03

The clearest mechanism is the sales loop. A prospective customer expresses interest in a retreat; Zo texts the business owner the person’s number; she calls while intent is still fresh; she asks Zo for a payment link; and she closes the booking. The website is only the visible entry point. Behind it, the server coordinates the lead notification, human conversation, payment step, and retreat records. Guo says the retreat-business user is “on track to make $100,000 on Zo”—a revenue projection, not achieved revenue. He also reports that the sales loop produces more booking revenue and enough demand to make retreats difficult to book, but the supplied recording provides no baseline, attribution analysis, or controlled comparison. 7:09

What must happen between initial interest and a completed booking? The flow below makes the division of labor visible: Zo moves information and prepares the transaction, while the owner handles the high-value conversation.

The retreat system contains a database even though its owner does not need to understand databases. It also keeps notes and accounting, and it can move retreat images onto the website when asked. This is the core abstraction: the user expresses an operational change in ordinary language, while the agent edits the files or data and serves the result from the user’s server. 8:33

Zo tries to span two audiences. Nontechnical users can ask it to build personal software, while developers receive a configured Linux virtual machine with root access, networking, SSH, API control, and MCP access. It can host personal agents and arbitrary services. That power is also a responsibility boundary: root access and arbitrary HTTP or TCP services expand what users can build, but they also expand the consequences of insecure software or over-permissioned agent actions, safeguards the talk does not detail. 9:32

The workspace demo groups several parallel capabilities around the same server:

• Models and agentic file work. Users can bring an API key, chat with different models, and let an agent work against files stored on the server.
• Storage and automation. The file system doubles as cloud storage, while scheduled tasks can run AI work automatically.
• Integrations and browser use. Built-in integrations, skills, and a logged-in browser let the agent operate external sites; Guo jokes that texting Zo to buy things makes impulse purchasing dangerously easy.
• Hosting. The server can expose arbitrary HTTP or TCP services and includes a personal website editable through chat.
• Personal software. Guo’s scheduling replacement applies his preferred rules, including reviewing people before allowing them to book time directly.

These features matter together. A chatbot alone can suggest a scheduling policy; a server-backed agent can store the policy, present a booking interface, evaluate incoming requests, and host the resulting service. The value comes from closing the loop between conversation and persistent software. 10:33

The closing argument widens from personal software to the structure of the internet. Companies already have domains, servers, APIs, and durable organizational presence. Guo wants individuals to have comparable homes from which they can interact directly with other people, companies, and company AIs rather than always passing through intermediary applications. 12:02

Agents make the ownership question urgent because many of them will run in the cloud. The user may experience a conversational assistant, but that assistant acts through conventional websites, APIs, files, and services, and its state must live somewhere. Some agents can remain local; many always-on agents will need remote computation and availability. “Whose cloud is it?” therefore determines who controls the agent’s context, operating environment, and accumulated improvements. 12:20

Guo points to a company-level cloud agent as an ergonomic pattern: many employees can interact with an assistant carrying company context. His objection is directional. If the organization spends time supplying context and refining behavior inside a provider-owned system, he argues that the provider captures the compounding improvement. He calls this intelligence feudalism: the organization uses the intelligence, but the intelligence “bubbles up” somewhere else. 12:52

Which direction does improvement travel? This comparison exposes the difference between renting an agent and publishing one on infrastructure the individual or company controls.

Zo’s proposed future lets an individual or company publish an agent, serve end users, and have those interactions improve the agent they own. This extends the personal-server idea into a software-company model: the owner controls the environment, the service, and the learning loop rather than only subscribing to somebody else’s application. Guo says Zo has a beta version of this agent-publishing platform. 14:05

That ending sharpens the talk’s title. “Everyone gets a software company” does not just mean everyone can generate a website. It means an individual can own a persistent server, combine data and AI there, turn repeated work into hosted software, and eventually publish agents whose improvement flows back toward their owner. The unresolved work lies in making that ownership operationally real: safe permissions, reliable deployment, backups, portability, evaluation, and clear control over what an agent learns remain necessary even when the interface becomes plain language. 14:05