OpenClaw Proves That AI Agents Don’t Need to Be Vertically Integrated — And Big Tech Hates It
One Austrian developer built an open-source AI agent that lets you swap between Claude, ChatGPT, and DeepSeek with a single command. It sparked a wave of competitors, triggered a regulatory conversation, and exposed a fundamental design choice Big Tech is quietly making — at your expense.
Google Gemini can now order your food through your phone. Microsoft Copilot can build your PowerPoint, read your email, and schedule your meetings. AI agents, it seems, are the next extension of existing monopolies — new tentacles for the same few platforms that already control your search, your documents, and your operating system.
— The core tension that OpenClaw quietly resolvedWhat Is OpenClaw? The Project That Surprised Everyone
OpenClaw is an open-source AI agent built by Peter Steinberger, an Austrian developer. What makes it remarkable is not what it does — it is how it does it.
Unlike the AI agents being rolled out by Google, Microsoft, and OpenAI, OpenClaw is not tied to any single foundation model. You can run it on Claude, ChatGPT, DeepSeek, or any open-source alternative. Switching between these options requires nothing more than a one-line command typed into the interface.
That single capability — frictionless model switching — sounds like a technical convenience. In reality, it is a political statement. It challenges the assumption that AI agents must be built, controlled, and owned by the same companies that build the underlying AI models.
These products reveal something important: AI agents are not a monolith. The foundation model and the software layer that surrounds it are in fact separable. Vertical integration is not a requirement.
— Pankaj Dubey, Brands Awareness, May 19, 2026The Gateway: How OpenClaw Actually Works
At the heart of OpenClaw is a component called the Gateway. Understanding the Gateway is the key to understanding why this project is architecturally different — and why that difference matters enormously for users.
⚙️ OpenClaw Architecture — The Gateway Model
📅 Calendar
📁 Files
Stores memory locally
Manages all connections
DeepSeek / Nemotron
Any open model
The Gateway is the key innovation. It sits between your data and the AI model — managing connections to your email, calendar, and apps, storing your preferences and memories locally on your machine. When you give it a task, the Gateway feeds only the relevant context to whichever model you are currently using. Because your data lives locally and connections are managed by the Gateway, switching foundation models does not mean starting over. Your integrations and memories follow you.
In practical terms: if you ask OpenClaw to book a restaurant, the Gateway pulls your calendar availability and dietary preferences, feeds them to the AI model of your choice, and completes the task — without that model ever holding a permanent record of your data.
This is the fundamental contrast with how Google, Microsoft, and OpenAI have designed their agents. In those systems, the company controls both the model and the data layer. With OpenClaw, the data layer belongs to the user.
Why This Architecture Matters: Modular vs Monolithic
To understand why OpenClaw’s design is significant, you need to understand what the alternative looks like — and why tech platforms prefer it.
The OpenClaw Wave: Every Major Copycat That Followed
Perhaps the clearest signal of OpenClaw’s significance is the response it provoked. The project prompted nearly every major Chinese tech company to launch an equivalent product — a rare real-time validation that the concept had struck a nerve.
OpenClaw Original
Peter Steinberger (Austria)
The original. Open-source AI agent with local Gateway, model-agnostic, fully portable memory and integrations.
Miclaw China
Xiaomi
Xiaomi’s OpenClaw equivalent — integrated with MIUI and Xiaomi’s device ecosystem.
Kimi Claw China
Moonshot AI
Kimi’s agent offering — building on the modular agent design principles of the original.
AutoClaw China
Zhipu AI
Zhipu AI’s answer to the growing modular agent market, targeting enterprise automation use cases.
NemoClaw Notable
Nvidia
Built directly on OpenClaw’s framework. Defaults to Nvidia’s Nemotron models. Added security and privacy tools. Also supports model swapping.
Nvidia’s entry is particularly significant. A hardware giant — not traditionally an AI agent company — built a product directly on OpenClaw’s open framework, validated its security model, and shipped it with enterprise-grade privacy tools. This is what open-source leverage looks like in practice.
The Problem with Vertically Integrated AI Agents
To understand why OpenClaw matters, you have to understand what the leading AI companies are actually building — and what that means for users over time.
Google’s Gemini now appears across Android, Google Search, and the entire Workspace suite. It can search through your photos and read your group chats. Microsoft has bundled Copilot with Office 365, with Agent Mode now integrated across Teams, Outlook, and Word. OpenAI’s ChatGPT can shop for you, connect to your health apps, and access your medical records.
| Platform | What Their Agent Can Access | Integration Model |
|---|---|---|
| Google Gemini | Photos, group chats, Gmail, Calendar, Android apps, Google Search, Workspace files | Vertical — bundled into Android + Google Workspace |
| Microsoft Copilot | Word, Excel, Outlook, Teams, SharePoint, PowerPoint, third-party connectors | Vertical — bundled into Office 365 |
| ChatGPT (OpenAI) | Shopping, health apps, medical records, memory, finance accounts via Plaid | Expanding vertically — now adding ads |
| OpenClaw | Email, calendar — plus any service user chooses to connect via Gateway | Modular — user controls all connections locally |
Each new agentic feature individually seems valuable. The compound picture is more troubling. An AI agent that has access to your location, your shopping history, your medical records, your group chats, and your calendar — managed by a single centralised company — is not just an assistant. It is the most detailed behavioral profile ever assembled on a human being, held by an actor with strong financial incentives to exploit it.
Surveillance, Lock-In & Self-Preferencing: The Three Compounding Problems
- Surveillance at scale: Assembling this kind of behavioral profile has traditionally required data brokers and tracking cookies working over years. With agentic AI, the information flows directly and continuously to one centralised actor. OpenAI has already moved to monetise this by introducing ads to ChatGPT.
- Self-preferencing: Any company that both develops and distributes an agent has obvious incentives to steer users toward its own products or those of paying commercial partners — less visibly than a search engine, since an agent typically presents a single recommendation with no page of alternatives.
- Lock-in through memory: Over time, an AI agent learns your habits, preferences, and context. This knowledge — both what you told it and what it inferred — is not easily transferred. Starting over means reconstructing this context and reconnecting all your apps and files. With each passing day, switching becomes harder.
Where a search engine at least shows a page of alternatives, an agent often presents a single recommendation, with little insight into how that recommendation was reached.
— Pankaj Dubey, Brands AwarenessThe self-preferencing risk is not theoretical. Google was fined €2.4 billion by the European Commission for burying rival shopping services beneath its own in search results. An AI agent can achieve the same outcome — steering you toward Google Shopping, Microsoft Azure services, or OpenAI partner brands — with far less visibility. There is no page of ten blue links to audit.
How OpenClaw Solves All Three Problems
OpenClaw does not eliminate these risks by accident. Its architecture directly addresses all three through deliberate design choices that prioritise user sovereignty over platform convenience.
| Problem | How OpenClaw Addresses It |
|---|---|
| Surveillance & Data Centralisation | Memories and activity logs are stored directly on your device in human-readable, editable files. No single actor accumulates a complete record. Users can rotate AI model providers — fragmenting any behavioral profile. |
| Lock-In Through Memory | Because data lives in the Gateway (local), switching AI models carries no memory cost. All integrations — email, calendar, preferences — persist across model changes. Portability is structural, not optional. |
| Self-Preferencing by Platform | OpenClaw has no commercial incentive to steer users toward any product. It has no advertising model, no ecosystem to protect. The agent recommends what the user’s chosen model and local context suggest — nothing more. |
| Sensitive Task Privacy | Users can switch to smaller open-weight models that run fully locally — so that not even the AI model provider sees the data. Medical questions, financial planning, and personal communication can stay entirely on-device. |
Lock-in is a design choice, not a technical reality. OpenClaw shows that memories and integrations can persist across model switches — making switching frictionless by design.
— Pankaj Dubey, Brands AwarenessThe Risks OpenClaw Introduces — Honest Assessment
OpenClaw is not a perfect solution. Intellectual honesty requires acknowledging the new risks that modular, open architectures introduce — and they are real.
- Malicious add-ons: Security researchers found 341 malicious add-ons on ClawHub, the add-on marketplace for OpenClaw extensions. These were designed to steal user data from the very Gateway that was supposed to protect it.
- Prompt injection attacks: A hidden instruction embedded in an email subject line, a webpage, or a document can hijack the agent’s behaviour — redirecting its actions to serve an attacker’s agenda rather than yours. An AI agent that reads your email is also an AI agent that can be manipulated through your email.
- Distributed security responsibility: With a vertically integrated agent, one centralized provider handles security for the entire system. With modular designs, security responsibilities are distributed across the user, the Gateway, the chosen AI model, and any third-party add-ons — a more complex threat surface to manage.
- User burden: OpenClaw puts meaningful control in the hands of users — but also meaningful responsibility. Non-technical users may struggle to maintain the security hygiene that a modular system demands.
These are not reasons to dismiss OpenClaw’s approach. They are reasons to build better infrastructure around it. And that infrastructure, as Jack FitzGerald notes in the source analysis, is already beginning to emerge — with Nvidia’s NemoClaw and a wave of startups building security and privacy layers specifically for open agentic systems.
The Regulatory Dimension: Can Modular Win Without Policy?
Here is the uncomfortable truth that OpenClaw’s existence surfaces: technical elegance is not enough. Without regulatory intervention, modular agent designs face a fundamental market disadvantage.
⚖️ The Regulatory Reality
Google and Microsoft have embedded their agents into their dominant software products. Gemini is in Android, in Search, and across Workspace. Copilot is in Office 365 — used by 400 million people. These integrations make vertically integrated agents effectively unavoidable for most users. A modular alternative cannot compete on distribution when its rivals are pre-installed.
Meta demonstrated the same dynamic when it blocked rival AI assistants from WhatsApp in favour of its own Meta AI. The Italian Competition Authority and European Commission intervened quickly with interim measures — but the key word is “unusually visible.” Platforms that never open up in the first place achieve the same outcome without triggering scrutiny.
Ultimately: unless regulators require gatekeepers to provide meaningful interoperability and data portability, projects like OpenClaw cannot offer a viable mainstream alternative — however superior their design philosophy.
| Regulatory Action | Against Whom | Outcome |
|---|---|---|
| €2.4B fine for self-preferencing | Google (EU Commission) | Google Shopping rivals were buried in search results — template for AI agents |
| Interim measures on WhatsApp | Meta | Italian CA + EU Commission blocked Meta’s ban on rival AI assistants |
| DMA interoperability requirements | EU Designated Gatekeepers | Ongoing — provides framework to mandate AI agent portability |
| Data portability standards | Policy proposal (not yet enacted) | Analysts argue users should have right to export memories, chat histories, and files in standardized formats |
The Irony: The Creator of OpenClaw Joined OpenAI
“The Austrian Developer Who Challenged OpenAI — Then Joined Them”
In February 2026, Peter Steinberger — the developer who built OpenClaw and, by doing so, demonstrated that one person’s open-source project could challenge the AI agent model being pursued by the world’s most valuable companies — was hired by OpenAI. Sam Altman himself announced the hire. OpenClaw remains open source. But its creator now works for the platform whose approach his project most directly challenged. Make of that what you will.
This twist does not invalidate OpenClaw’s contribution. The project, the code, the architecture, and the ideas it has seeded across the industry all persist. NemoClaw, Miclaw, Kimi Claw, and AutoClaw are all running. The blueprint for agent portability is written, published, and freely available to anyone who wants to build on it.
What a Modular AI Future Looks Like — And What It Needs
OpenClaw points toward a future that is meaningfully better for users — if the right conditions are in place. Here is what that future requires:
- Data portability as a right: Users should be able to export their AI agent’s memories, chat histories, uploaded files, and learned preferences in standardised, human-readable formats — and move them to a competing agent without data loss.
- Interoperability mandates for gatekeepers: Platforms like WhatsApp, Android, and iOS should be required to allow competing AI agents access to their APIs and core platform features — not just the platform’s own agent.
- Security infrastructure for open agents: The modular ecosystem needs shared security standards, vetting systems for add-on marketplaces, and prompt injection safeguards that work across agent implementations.
- Meaningful choice architecture: Users should be able to choose their AI model at the agent layer — not be locked in by a device pre-install or a software bundle. Competition at the model layer requires portability at the agent layer.
The good news: the market is already moving in this direction. OpenClaw exists. NemoClaw exists. The blueprint is public. The question is whether platform power, distribution advantages, and regulatory inertia will bury that blueprint — or whether policy will give it the conditions to compete.
Frequently Asked Questions
The Blueprint Exists. The Choice Is Whether to Follow It.
OpenClaw did not just build an AI agent. It built proof-of-concept for a different future — one where users own their data, choose their AI model, and are not held hostage by the platform that happens to be pre-installed on their phone or bundled with their office software.
That future is technically possible. Peter Steinberger demonstrated it alone, in his spare time. Nvidia, Xiaomi, Moonshot AI, and Zhipu all confirmed it by building on the same blueprint within months. The question that remains — and the question that regulators must answer — is whether Big Tech’s distribution advantages will make that blueprint irrelevant, or whether policy will give modular, open AI agents the conditions they need to genuinely compete.
