Anthropic this week revised its legal terms to clarify its policy prohibiting the use of third-party harnesses with Claude subscriptions, as the AI industry looks to strengthen its revenue model.
Anthropic sells subscriptions to the Claude platform. This provides access to a family of machine learning models (such as Opus 4.6) and related tools such as Claude Code, Claude.ai’s web-based interface, and the Claude Desktop application.
Claude Code is a harness or wrapper. It integrates with the user’s terminal and works with other tools and control loops to route prompts to available Claude models. These loops make it what Anthropic calls an agent coding tool.
Many other tools act as harnesses for the model, including OpenAI Codex, Google Antigravity, Manus (recently acquired by Meta), OpenCode, Cursor, and Pi (the harness behind OpenClaw).
The harness exists because interacting with a machine learning model is not itself a good user experience. Enter the prompt and the results will be returned. It’s a one-turn interaction. input/output. To create products that people care about, model makers have added support for multi-turn interactions, remembering previous interactions, accessing tools, and orchestrating data flowing between tools. Some of this support is built into the model platform, while others are added through the harness tool.
This can create business problems for advanced model manufacturers. They invest billions of dollars to train sophisticated models, but risk being disintermediated by gatekeeping intermediaries who build harnesses around the models and provide a better user experience.
One of the ways Anthropic has chosen to build brand loyalty is by selling tokens to subscription customers at a monthly price with limited usage. This makes it cheaper than buying tokens on a pay-as-you-go basis via the Claude API. Essentially, the economics are similar to an all-you-can-eat buffet where prices are set based on specific usage expectations.
This practice was effectively a subsidy to subscribers and led to token arbitrage. Customers accessed the Claude model through a subscription linked to a third-party harness because it cost less than doing the same work through an API key.
AI Business’ Consumer Terms of Service prohibits the use of third-party harnesses without specific permission starting at least February 2024. The contract language in Section 3.7, which has not changed since then, says as much, prohibiting unauthorized automated access tools.
You may not access or use, or assist anyone else in accessing or using, our Services in any way that:
However, unless you are accessing our Services through an Anthropic API key or as expressly authorized by us, you will access our Services through automated or non-human means, such as through bots, scripts, etc.
Although this provision has existed for more than two years, various third-party tools have ignored this rule and allowed users to enter their cloud subscription account keys.
The rules added clarify that OAuth authentication, which is the access method used for Claude Free, Pro, and Max level subscribers, is only for Claude Code and Claude.ai (the web interface for Claude models).
”“Using OAuth tokens obtained through Claude Free, Pro, or Max accounts with any other product, tool, or service, including the Agent SDK, is not permitted and is a violation of our Consumer Terms of Service,” the updated legal compliance page states.
Anthropic says the update represents an attempt to clarify the language of existing policies and make them consistent across company documents.
It appears Anthropic decided to crack down on that rule earlier this year. Anthropic engineer Thariq Shihipar said in a social media thread in January that the company had taken steps to prevent third-party tools from “spoofing Claude Code Harness.”
“Third-party harnesses that use Claude subscriptions cause problems for our users and are prohibited by our terms of service,” he wrote. “They generate unusual traffic patterns without the normal telemetry that the Claude Code harness provides, making it very difficult to help debug when there are questions about rate limiting usage or account bans and there is no other means to provide this support.”
The ban proved unpopular enough to elicit a reaction from the competition. OpenAI’s Thibault Sottiaux was a vocal supporter of using Codex subscriptions with third-party harnesses.
After Anthropic banned accounts that attempted to cheat its pricing structure, which Shihipar suggested is coming, third-party harness manufacturers are also taking note.
On Thursday, OpenCode pushed code that removes support for Claude Pro and Max account keys and Claude API keys. This commitment cites “the legal imperatives of humanity.” ®
