Anthropic Officially (More Clearly) Bans 3rd-Party Harnesses with Claude Code API :-(

Claude Code usage is governed by Anthropic's legal terms, with specific provisions for consumer, commercial, and healthcare compliance. The policy mandates strict authentication standards, requiring API keys for developers while limiting OAuth tokens to individual consumer use. Security and trust are managed through dedicated transparency hubs and a formal vulnerability reporting system.

Key Points

• Usage is governed by specific terms based on the user's plan, with existing commercial agreements applying to both direct and third-party API access.
• Healthcare compliance through BAAs is supported for Claude Code if Zero Data Retention (ZDR) is enabled for the account.
• Strict authentication rules separate individual OAuth usage from developer API key requirements, prohibiting the use of consumer tokens in third-party tools.
• All users must adhere to the Anthropic Usage Policy, with individual plan limits assuming ordinary usage patterns.
• Security and transparency are maintained through the Anthropic Trust Center and a formal vulnerability reporting process on HackerOne.

Sentiment

The community is predominantly negative toward Anthropic's decision. While a substantial minority defends the business rationale of protecting a subsidized service, the majority views it as a hostile anti-competitive move that prioritizes lock-in over developer experience. The tone is frustrated rather than purely hostile, with many expressing disappointment in a company they previously supported. A recurring theme is that this will backfire as competitors like OpenAI offer more permissive terms.

In Agreement

• Subscriptions are heavily subsidized loss leaders — Anthropic has every right to restrict how that subsidy is consumed, just as a restaurant giving 90% off coupons can require in-store dining only
• The subscription explicitly pays for Claude Code as a product, not raw API access — comparing it to Netflix or PlayStation Plus, you pay for the service through their interface
• Third-party clients may lack the caching optimizations that make subscription pricing viable, potentially costing Anthropic significantly more per request
• Users wanting cheap unrestricted API access through subscription tokens are effectively exploiting a subsidy meant for a specific product experience
• The API with per-token pricing exists for anyone who wants complete freedom of client choice — the two products serve different needs

Opposed

• Users paying $100-200/month should be able to choose their preferred coding tool — they are not consuming beyond their quota, just using a different interface
• Claude Code is buggy, memory-hungry (reports of 55GB RAM usage), and inferior to alternatives like OpenCode and pi, making this a forced downgrade
• This is a textbook anti-competitive tie-in sale that is illegal in some jurisdictions like Brazil and echoes the Microsoft bundling controversies
• OpenAI takes the explicitly opposite approach by allowing subscription tokens in third-party tools, making Anthropic look hostile to developers
• This is classic enshittification — subsidize to build market share, then restrict freedom and lock users into an ecosystem
• Anthropic's moat is the model quality, not the harness — this move will accelerate user migration to competitors as models become increasingly commoditized
• Many users report creating workarounds like running Claude Code headless or through ACP protocol, suggesting the ban creates an unwinnable cat-and-mouse game