Anthropic revised its consumer terms in the week of February 16 to make explicit that Claude Free, Pro and Max subscriptions may not be used to power third-party tools such as OpenClaw, The Register reported on February 20. The company’s position is that the subsidized token allowances included in a monthly subscription were being used at scale through outside software, raising its costs and undercutting the economics of its pricing.
The change is less a new rule than a clarification. Anthropic’s terms already prohibited using OAuth credentials from a consumer account in unapproved tools. The updated wording states that subscription tokens are for Claude Code and Claude.ai only, and that routing them through any other product or service is a breach of the agreement.
What a harness is
Anthropic sells access to its models, including Claude Opus 4.6, through several surfaces: the Claude.ai web and desktop apps, the Claude Code agent for software development, and a pay-per-token API. Claude Code is what developers call a harness: a layer of software that connects a model to a user’s terminal, tools and control loops so that it can carry out multi-step tasks. A bare model takes a prompt and returns a response; the harness adds memory, tool use, multi-turn planning and automation.
Many independent harnesses exist, and some offer features or interfaces that users prefer. OpenClaw, an open-source agent framework, is one of the more popular. Because these tools can be pointed at any model provider, a user with a Claude subscription could log in through OAuth and run the third-party harness against Anthropic’s models at subscription rates.
The arbitrage problem
Subscriptions bundle a limited quantity of usage for a flat fee, priced on the assumption of typical use inside Anthropic’s own apps. Per-token API pricing is higher for heavy use. Running an aggressive agent harness on a subscription therefore delivers API-level consumption at a fraction of API cost, a practice critics and defenders alike have called token arbitrage. Anthropic says this was happening at a scale that mattered.
The company has been enforcing the rule since early in the year. An Anthropic engineer explained, according to The Register’s account, that third-party harnesses generate unusual traffic patterns and do not send the telemetry the company’s systems expect, which makes it harder to debug rate limits and account restrictions. Some accounts were suspended, prompting complaints from developers who said they had not understood the rule.
Disintermediation risk
Underneath the cost argument is a strategic one. Model developers have spent billions of dollars training frontier systems. If an outside company builds a better interface on top of those models and captures the customer relationship, the model maker risks being reduced to a commodity supplier. Keeping subscription users inside its own harness protects Anthropic’s direct relationship with them.
Rival providers have taken different stances. Some have openly welcomed third-party harnesses as a route to more usage, while some tool developers have removed Claude subscription login after receiving legal requests from Anthropic, according to the reports. Users who want to run Claude in an independent harness can still do so through the API, at API prices.
The dispute points to a larger question for the industry: whether power will rest with the companies that build foundation models or with the platforms that deliver the best experience around them. Tight control by model makers limits the open ecosystem; dominance by harnesses squeezes the model makers’ revenue. The AI race is increasingly about control, access and business models, not only about which model is smartest.
