WTF Anthropic

Anthropic has tightened access rules for its Claude Code “harness,” restricting paid API subscriptions so they can be used only with Claude Code—blocking third-party coding tools that previously worked by spoofing or “forging” requests. The change triggered a wave of backlash from developers who had relied on tools like Cursor and Open Code, with many scrambling to find workarounds and complaining that the enforcement arrived suddenly.

The controversy centers on a credential restriction: a token authorized for Claude Code cannot be used for other API requests. Until this week, some third-party tools reportedly managed to operate under Anthropic’s subscription terms by using the same underlying API token while sending requests with custom headers and payloads designed to mimic Claude Code’s expected traffic. That approach is described as fragile—dependent on undocumented behavior and prone to break when Anthropic changes safeguards. In practice, the terms of service for Claude Code had long discouraged this kind of API use outside Anthropic’s products, meaning the “surprise” may have been less about a new policy and more about enforcement finally catching up.

A related explanation comes from a claim attributed to the Claude Code operator on Twitter: Anthropic says it tightened safeguards against spoofing after accounts were banned for triggering abuse filters tied to third-party harnesses using Claude subscriptions. The operator also argues that third-party harnesses create unusual traffic patterns and lack the telemetry Claude Code provides, making rate-limit, usage, and account issues harder to debug. The practical effect is that when problems occur, developers may see them as Anthropic’s fault rather than the third-party tool’s—so Anthropic is now prioritizing consistent enforcement to protect supportability and reduce abuse.

Why enforce the restriction now? Cost is the most common theory, but the analysis here treats it as incomplete. The argument offered is that Anthropic’s real incentive is to make its entire stack “sticky”: if developers can only get the best token value through Claude Code, then Claude Code becomes the gateway product, and Anthropic captures more of the workflow value. That matters because Open Code is portrayed as rapidly growing toward one million monthly active users, and the speaker claims Claude Code’s usability has lagged—citing interface bugs like flickering and rendering glitches.

The transcript also advances a more speculative “tinfoil” view: Anthropic may be heavily subsidizing its higher-tier plans far beyond what pricing charts suggest, because model training and inference costs remain high and competitive pressure forces constant hardware upgrades. The claim is that new GPU generations and shifting model options create a continual arms race, so restricting third-party access helps preserve margins while steering users toward Anthropic-controlled tooling. The segment ends with a broader accusation that Anthropic leadership—named as Dario Amodei in the transcript—wants control over AI development and regulation, framing the enforcement as part of a larger effort to limit independence and keep developers dependent on proprietary systems.

Overall, the core takeaway is straightforward: Anthropic is moving from permissive “it works until it doesn’t” token usage to strict, terms-based enforcement, and the developer ecosystem is reacting as if a long-standing workaround has been cut off without warning—whether the timing reflects economics, product strategy, or both.