Can you use your Claude subscription with OpenClaw?
Anthropic banned it in April 2026, then walked it back twice. What its terms say now, why agent workloads exhaust subscription limits, and the predictable alternatives.
Mechanically, yes: OpenClaw can draw on your Claude subscription’s usage limits, in two ways. One has OpenClaw store your subscription token, which Anthropic’s terms bar third-party developers from doing. The other runs your own, unmodified Claude Code, which the terms let you sign in to; whether an agent driving it all day is the “ordinary use” they describe is Anthropic’s call, enforced “without prior notice” on your account. Anthropic banned this use in April 2026 and has moved three times since, and agent workloads chew through subscription limits fast. For a long-running agent, an API key (or OpenRouter, or a ChatGPT subscription) is the predictable path.
The 2026 whiplash, dated
If you’ve seen contradictory answers to this question, that’s because all of them were true at some point this year:
- April 4, 2026: Anthropic blocks Claude Pro and Max subscriptions from powering third-party agents like OpenClaw. The stated reason was blunt economics: subscribers paying $20–200/mo were burning far more than that in agent-driven usage, without the response caching that makes flat-rate chat viable.
- May 14, 2026: Anthropic announces “Agent SDK credits”: a fixed, non-rollover monthly credit for programmatic use, billed at API rates, to start June 15, which would have restored third-party agent access in a fenced-off form.
- June 15, 2026: a support update pauses that separate credit plan before it starts, so Agent SDK,
claude -pand third-party app usage keep drawing on the subscription’s ordinary limits. - October 7, 2026: a further update on the same support page says Max and Team plans “now include monthly API credits, which cover the Claude Agent SDK, claude -p, the Claude API, and Claude Managed Agents,” and that “You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits.”
Anyone building on this arrangement should assume it can move again: it has moved four times in six months.
How it works today, and where each route stands
OpenClaw’s Anthropic integration offers two subscription routes and an API-key route. The first stores a token you mint with claude setup-token on any machine and paste in on the gateway host:
openclaw models auth login --provider anthropic --method setup-tokenThe second, --method cli, runs the Claude Code installed on the gateway host, which keeps its own login and renews it.
After either, Claude’s models are selectable and usage counts against your plan’s limits. The API-key route (openclaw onboard with your Anthropic key) bills per token with no plan entanglement. OpenClaw’s own docs make the recommendation plainly: for production automation and predictable costs, use the API key.
Anthropic’s terms for Claude Code, as of October 3, 2026, bear on the two subscription routes differently. They bar third-party developers from routing requests through Pro or Max credentials for their users, and say developers “may not collect, store, or intermediate Claude.ai credentials or session tokens.” The setup-token route is that shape: OpenClaw stores your token and sends requests with it itself. The CLI route, as long as no token is saved in OpenClaw, leaves the login inside Claude Code, run as Anthropic ships it, and the same page does not prevent “an end user from signing in to the unmodified Claude Code binary with their own Claude subscription.” Anthropic’s help center says third-party app usage draws from your plan’s limits. The terms also describe subscription sign-in as “designed to support ordinary use of Claude Code and other native Anthropic applications,” and whether an agent driving Claude Code all day is that ordinary use is Anthropic’s call.
Why the subscription path disappoints agent operators
A personal agent is not a chat session. Every turn re-sends workspace files, memory, and skill context (operators routinely see five-figure token counts per message), and an always-on agent takes many turns a day. That’s exactly the usage shape that made Anthropic act in April, and it’s why the current arrangement, where agent traffic shares your subscription’s limits, mostly means hitting those limits mid-week and falling back to metered billing anyway. Community write-ups of the restored access reached the same conclusion: it’s a nicer on-ramp to paying API rates, not genuine flat-rate coverage for a serious agent.
What we’d actually do
Treat Claude access as a model choice, not a billing plan: connect Claude by API key (directly or via OpenRouter) when you want Claude’s models for specific work, and power the agent’s daily traffic with access whose economics were designed for agents, which today means a ChatGPT subscription (OpenAI took the opposite policy path and officially supports it) or straightforward metered usage with a clear picture of what that costs. The deeper lesson of the April block is worth keeping: an agent you own should never depend on one provider’s subscription policy, which is precisely the argument for running an open-source agent that can switch brains without losing itself.