Anthropic says OpenClaw-style Claude CLI usage is allowed again

Original Article ↗ Hacker News Discussion ↗

Anthropic’s shifting stance on whether Claude subscription credentials can be used with third‑party CLI tools like OpenClaw has left many developers confused and frustrated. Commenters describe unclear, frequently changing rules around “harnesses,” surprise account actions, and opaque pricing/usage limits as eroding trust and pushing them toward competitors such as OpenAI’s Codex or Chinese models. Several argue Anthropic must clearly separate policies for consumer subscriptions vs. API usage and publish stable, transparent limits if it wants to remain a credible platform for serious tooling.

Policy confusion around CLI and OpenClaw

  • Multiple commenters find Anthropic’s stance on Claude CLI / OpenClaw use opaque and shifting.
  • Distinction discussed between:
    • Using OAuth credentials directly against Anthropic’s API or third‑party harnesses (said to incur “extra usage” and previously risk bans).
    • Running tools inside Claude Code CLI via claude -p (described as allowed in principle, but at times blocked by classifiers and/or billed as extra usage).
  • OpenClaw maintainers report Anthropic staff verbally OK’ing “CLI-style” usage, yet parts of OpenClaw’s system prompts were still blocked until recently, suggesting misalignment between public statements, classifiers, and ToS.
  • Several note this announcement comes via OpenClaw, not an official Anthropic channel, and consider it unreliable or “hearsay.”

Trust, reliability, and communication

  • Many say repeated changes to limits, policies, and unofficial “clarifications” (often via scattered tweets) have eroded trust.
  • Some accounts were disabled for “suspicious signals” with no clear recourse; others later found their accounts silently re‑enabled.
  • Users fear future rug pulls and see the environment as too unstable for critical workflows.

Pricing, limits, and economics

  • Complaints about reduced subscription limits, rising prices, and the opaque split between “plan limits” and “extra usage.”
  • Some view CLI/OpenClaw billing as effectively “API with extra steps.”
  • Debate on whether LLM providers are currently losing money on inference and how sustainable low consumer prices are.

Harnesses vs models / product strategy

  • Strong sentiment that the “harness” (Claude Code, Codex, OpenClaw, pi, etc.) is where much of the value lies.
  • Tension identified: Anthropic seems torn between maximizing Claude Code usage vs being a neutral model platform, leading to policy whiplash.
  • Several argue Anthropic should publish clear developer policies and simple rate limits tied to OAuth, rather than vague, case‑by‑case enforcement.

Switching behavior and model comparisons

  • Numerous users report canceling Claude plans (Pro, Max, 20x, 5x) and moving to OpenAI Codex, Chinese models (GLM, Z.ai, MiniMax, Kimi), or local models.
  • Opinions on model quality are split: some find Anthropic better at clean, readable code; others say GPT‑5.3/5.4 Codex is consistently stronger for implementation and planning.
  • Many say frontier models are now close in quality; harness UX and predictable pricing drive provider choice more than raw model differences.

Workarounds and setups

  • People describe using claude -p scripts, tmux, Telegram bots, MCP Channels, local‑cloud hybrids, and custom agents to stay within perceived rules while maximizing subscription value.
  • It remains unclear which of these patterns Anthropic fully endorses long‑term.