
Anthropic Is Secretly Building an Always-On AI Agent Called Conway. It Looks Like an OpenClaw Killer.
Anthropic cuts off third-party tools from Claude subscriptions. Days later, code reveals it is building Conway, its own always-on agent platform with extensions, webhooks, and Chrome integration.
The AI Post newsroom — delivering AI news at the speed of intelligence.
Last week, Anthropic told every developer using Claude through third-party frameworks like OpenClaw to pay up or get out. As of April 4, flat-rate Claude Pro and Max subscriptions no longer work through external tools. Developers face cost increases of up to 50x.
The timing was suspicious. Now we know why.
Testing Catalog has uncovered code revealing Anthropic is internally testing an always-on AI agent called Conway. It is not a chat window. It is not a coding assistant. It is a persistent agent environment with its own extension system, webhook triggers, Chrome browser integration, and a dedicated management interface. Think of it as Anthropic building its own OpenClaw, but with native Claude integration baked into every layer.
The leaked interface shows three core areas: Search, Chat, and System. Under System, there is an Extensions section where users can install custom tools, UI tabs, and context handlers using .cnw.zip files. That is a new extension package format nobody has seen before. There are also Connectors and Tools sections showing connected clients, and a toggle that lets Claude in Chrome connect directly to a Conway instance.
The most telling detail: a Webhooks section with public URLs that can wake the Conway instance when external services call them. This is not a tool you open and close. This is an agent that stays alive, listens for events, and acts on them. Always on. Always connected.
References to the previously spotted Epitaxy UI appear inside Conway, suggesting Epitaxy may serve as the operator interface for the whole environment. If Anthropic ships anything close to what the code reveals, it would be the company's biggest product launch since Claude itself.
Here is what makes this story matter beyond product news: the sequence of events tells you everything about Anthropic's strategy. Step one, cut off third-party agent frameworks from flat-rate pricing. Step two, build your own agent platform. Step three, own the entire stack from model to interface to extension ecosystem.
This is the classic platform owner playbook. Apple did it with apps. Google did it with search. Amazon did it with marketplace sellers. First you let the ecosystem build on your platform. Then you study what works. Then you build it yourself and change the economics so the original builders cannot compete.
The developers who built businesses on Claude through OpenClaw are now facing a choice: pay 50x more for API access, or wait for Anthropic to ship the tool that replaces the one they just broke. Neither option is great.
Anthropic has not publicly commented on Conway. Given that the code was found in testing environments, a public launch could be weeks or months away. But the direction is clear: Anthropic does not want to be just a model provider. It wants to be the platform. And it is willing to burn its developer ecosystem to get there.