Anthropic officially bans using subscription auth for third party use

Policy Change: What’s Now Banned

  • OAuth tokens from Free/Pro/Max subscriptions may only be used in Claude.ai and Claude Code.
  • Using those tokens in any third‑party product, tool, or service — explicitly including the Agent SDK — violates the Consumer ToS.
  • Third‑party apps must use metered API keys from Console or cloud providers; no “log in with Claude”–style flows for routing user traffic through subscriptions.
  • Anthropic says it can enforce this without notice; some users/tools have already been blocked or banned.

Targets and Affected Ecosystem

  • Clearly aimed at tools like OpenClaw / OpenCode‑with‑Claude and similar agentic coding harnesses that authenticated via Claude subscriptions (often by spoofing Claude Code).
  • Any app that uses the Agent SDK with subscription OAuth (including personal projects) is, per the written docs, out of bounds.
  • Wrappers that only shell out to the official claude CLI or Claude Code binary (e.g. simple TUI/GUI shells or ACP clients) are generally seen as still acceptable, though edge cases (e.g. modified Claude Code binaries) have been blocked.

Confusion Around the Agent SDK and Messaging

  • The docs say subscription OAuth cannot be used “in any other product, tool, or service — including the Agent SDK.”
  • A product leader on X claimed “no changes” to how SDK and Max work and suggested personal experimentation is fine, contradicting the ToS.
  • Many commenters dismiss tweets as non‑binding compared to the contract; others see the mismatch as a PR and legal risk.

Economics and Motives

  • Widely shared view: flat‑rate subs are heavily subsidized and priced far below equivalent API usage, especially for Max; power users can burn thousands of dollars of tokens for $200/month.
  • Third‑party agents can max out weekly/5‑hour quotas automatically, destroying the assumed “human‑paced” usage Anthropic priced for.
  • Some argue limits alone should suffice; others note caching, usage patterns, and arbitrage make third‑party harnesses uniquely costly.

Lock‑in, Alternatives, and Backlash

  • Many see this as a deliberate walled‑garden move: tying cheap subscriptions to mediocre first‑party tools (especially Claude Code), blocking better open harnesses.
  • Comparisons are drawn to gym memberships, “enshittification,” and Apple‑style ecosystems.
  • Several users report cancelling or downgrading Claude and moving to OpenAI Codex, Kimi/GLM, MiniMax, DeepSeek, Gemini, Mistral, or local models; Codex’s explicit support for third‑party harnesses is repeatedly cited as more developer‑friendly.