I Analyzed 512,000 Lines of Leaked Code. It Shows What's Coming for Your AI Tools.
TL;DR
The big story in Anthropic’s leak is Conway, not the code — buried in 512,000 leaked lines was an unannounced “always-on” agent environment with browser control, external triggers, extensions, and tool connectors that looks less like chat and more like an agent operating system.
Anthropic is executing a five-part platform strategy at speed — Nate ties Conway to Claude Code channels, Claude Co-Work, Claude Marketplace, partner-network investments, and new restrictions on third-party access, arguing this is a coordinated lock-in play, not a set of random launches.
Conway’s proprietary extension layer may matter more than MCP’s openness — while Anthropic helped popularize the open Model Context Protocol, Conway’s CNW.zip format adds a Conway-only app layer, which Nate compares to Google Play Services sitting on top of open Android.
The real lock-in is behavioral memory, not your files — unlike Salesforce records or Slack history, a persistent agent learns how you work over months, and there’s no CSV export for “what this person prioritizes, ignores, or drafts well.”
The OpenClaw crackdown looks like a familiar platform pattern — Nate cites Peter Steinberger joining OpenAI on February 14, Anthropic’s January-to-February policy tightening, and a later enforcement wave as evidence of a sequence: copy the feature, subsidize the first-party version, then make third-party access expensive or impossible.
2026’s AI battle is shifting from model quality to persistent agent ownership — Nate’s core claim is that Google, OpenAI, and Anthropic now treat models as loss leaders and are racing to own the memory layer that stores your workflows, context, and organizational knowledge.
The Breakdown
The leak’s real bombshell: Conway
Nate opens by saying the most important thing in Anthropic’s leak “isn’t the code” but an internal project called Conway. In the leaked source, Conway shows up as a standalone agent environment with its own page, sidebar, extensions, browser control, tool connections, and external wake-up triggers — not just another Claude chat tab.
What your morning looks like when an agent never sleeps
He makes it vivid with a Tuesday-morning scenario: Conway watches email, Slack, competitive intel, dashboards, and your calendar overnight, then drafts responses and board-prep context before you type a word. The catch is crucial: maybe a third of it is wrong, but if it’s fast and mostly useful, the net gain is still positive — which is why he says the value comes from speed and iteration, not perfection.
Conway only makes sense as part of Anthropic’s larger push
Nate zooms out to Anthropic’s last 90 days: Claude Code channels in Discord and Telegram, Claude Co-Work for non-engineers, Claude Marketplace for enterprise procurement, and a reported $100 million commitment to the Claude partner network, with Accenture training 30,000 professionals. Layer onto that Anthropic’s move to block third-party tools from Claude subscriptions, and he says you’re looking at a unified harness strategy across developer tools, enterprise software, distribution, and enforcement.
The Microsoft-in-the-’90s analogy
This is where he goes big: Anthropic isn’t just shipping features, it’s “speedrunning” the Microsoft arc from DOS to Windows to Office to Active Directory. In his framing, Conway is the Active Directory move — the sticky layer that knows your organization deeply enough to make the rest of the stack hard to leave.
MCP is open — Conway’s extension model isn’t
One of the sharpest points in the video is about standards. Anthropic helped launch MCP as an open protocol adopted by OpenAI, Google, and hosted by the Linux Foundation, but Conway’s CNW.zip extension format creates a proprietary layer for UI panels, handlers, and tools that only work inside Conway. Nate compares it to Google Play Services: the open foundation gets goodwill, while the closed layer captures the value.
Why developers may follow the app-store gravity
He frames the developer choice in painfully familiar terms: build a portable MCP tool with no built-in distribution, or build a Conway extension that gets discovered inside Anthropic’s own environment. His analogy is early iPhone versus the open web — architecturally, the open path may be cleaner, but the money and attention tend to flow to the app store.
OpenClaw, Peter Steinberger, and the lockout pattern
Nate connects Conway to the recent OpenClaw drama. He notes that creator Peter Steinberger joined OpenAI on February 14, OpenClaw moved to a foundation with OpenAI backing, and Anthropic’s restrictions escalated from quiet blocking in January to revised terms in February to hard enforcement later. His read is blunt: first the platform copies popular community features into the first-party product, then it closes the door on open alternatives.
The next battleground is your behavioral model
The closing argument is the deepest one: previous lock-in was about files, records, and message history; Conway-style lock-in is about the accumulated model of how you work. Nate argues the industry needs a conversation — and ideally policy — around “behavioral context portability” before products like this fully launch, because once your productivity is bound up with a persistent agent, switching providers or employers could mean leaving part of your professional operating model behind.