Claude Code refuses requests or charges extra if your commits mention "OpenClaw"

Original Article ↗ Hacker News Discussion ↗

Anthropic’s Claude Code tool is reportedly throttling sessions and switching users to more expensive “extra usage” billing when their Git commits or files contain strings like “OpenClaw” or “HERMES.md,” even if they are not actually using those agent frameworks. Commenters see this as a mix of clumsy anti‑abuse measures against subscription‑draining agents and potentially anti‑competitive behavior, exacerbated by opaque limits, billing surprises, and quality regressions. The episode is driving many developers to explore alternatives such as OpenCode, DeepSeek, Kimi, and local open‑weight models, and raising broader concerns about reliability, trust, and lock‑in with closed AI platforms.

Reported OpenClaw / HERMES Behavior

  • Multiple commenters reproduce that certain strings (e.g., openclaw in git commits or HERMES.md) can cause Claude Code sessions to:
    • Immediately disconnect.
    • Mark 5‑hour usage as 100%.
    • Or silently route requests to metered “extra usage”/API billing instead of subscription quota.
  • Others report only refusals or errors, not quota exhaustion, suggesting behavior may be inconsistent, A/B tested, or recently changed.
  • Several note Claude sometimes denies knowing what “OpenClaw” is or refuses to acknowledge it even when given the URL, which some view as “gaslighting.”

Anthropic’s Motives: Abuse Control vs. Anti‑Competitive

  • One side: This is framed as a necessary anti‑abuse measure against tools like OpenClaw that run always‑on agents and can burn huge amounts of subsidized tokens; flat‑rate plans can’t support 24/7 automated agents.
  • Other side: Critics argue this crosses into anticompetitive behavior, punishing mere mentions of a competing harness, and effectively changing pricing terms in opaque ways.

Engineering Quality & “Vibe‑Coding”

  • The detection appears to be naive string/regex matching over project history, which many see as sloppy and fragile.
  • This is linked to the recently leaked Claude Code codebase and claims that Anthropic increasingly has Claude generate large parts of its own tools with limited human review.
  • Some argue this pattern (HERMES, OpenClaw, context bugs) suggests poor QA and a culture of shipping LLM‑generated patches without robust engineering safeguards.

Business Model, Compute Constraints, and Subscriptions

  • Broad agreement that frontier models are compute‑constrained and heavily subsidized; subscription economics break when users automate heavy workloads.
  • Debate over whether Anthropic should:
    • Raise prices or move fully to metered billing.
    • Cap usage more explicitly.
    • Stop onboarding or restrict higher tiers until capacity improves.
  • Several see current behavior as “enshittification” of flat‑rate plans driven by investor pressure and competition (including cheaper Chinese/open‑weight models).

Trust, Transparency, and Ethics

  • Many say the main issue is opacity: secret keyword triggers, undocumented billing paths, weak support, and retroactive non‑refund policies for expired credits.
  • Broader skepticism surfaces about Anthropic’s “safe/ethical” branding given:
    • Military and DoD work.
    • Aggressive PR versus reported internal practices.
  • Others counter that all major labs are similarly pressured by scale, and Anthropic may still be “least bad.”

Alternatives and User Migration

  • Significant interest in switching to:
    • OpenCode, Codex, Cursor, and other harnesses.
    • Frontier‑adjacent models like DeepSeek, GLM, Kimi, Qwen.
    • Local/open‑weight models via Ollama, llama.cpp, etc.
  • Mixed reports on quality: some find open models close to Sonnet/Opus for many tasks; others say they remain clearly inferior, especially for complex, autonomous work.