Anthropic fixes Claude Code proxy false account flags

• Anthropic’s public line is that Claude Code can run other models through a proxy, since bcherny pointed users to LiteLLM in his proxy reply.
• The escalated block was treated as an account-classifier false positive: bcherny said the user was unblocked in his unblock reply, then said the account was good to go in his update.
• Anthropic denied the broader rumor that harness use itself triggers bans, with bcherny saying false positives are being reduced in his follow-up.
• The credential boundary still matters: bcherny’s classifier explanation in his main reply sits alongside Anthropic docs that allow gateways but restrict third-party use of Claude.ai subscription credentials.

Claude Code’s gateway docs spell out the ANTHROPIC_BASE_URL and credential path. LiteLLM’s non-Anthropic model tutorial goes further: Claude Code can call GPT-4o, Gemini, Azure OpenAI, Vertex AI, and other providers through a proxy that serves the Anthropic Messages API. Anthropic’s legal page keeps another boundary intact: third-party developers cannot offer Claude.ai login or route requests through Free, Pro, or Max credentials. The same week, Anthropic made auto mode the default in Claude Code, another classifier layer, with classifier overhead no longer billed on Pro, Max, and Team.

Boris Cherny, listed by Anthropic as Head of Claude Code, answered the core accusation directly: bcherny said in his classifier reply that Anthropic does not ban people for using harnesses with other models, and that the escalated case was “almost certainly” triggered by a different account classifier.

The fix then moved fast:

• The account was “Unblocked” in bcherny’s one-word update.
• The account was “good to go” in bcherny’s follow-up.
• Anthropic was “working on making sure this doesn’t happen again to others,” according to bcherny’s update.
• bcherny later wrote “Agree, we need to do better” in his reply to the thread.

Christmas come early for coding-agent nerds, but only after a false positive locked somebody out of the toolchain.

bcherny said in his proxy reply that “using our harnesses with other models is supported,” and named LiteLLM as the proxy path.

The official Claude Code gateway docs describe an LLM gateway as a proxy between Claude Code and a model provider. The setup uses two moving parts, according to Anthropic’s gateway connection guide:

• ANTHROPIC_BASE_URL points Claude Code at the gateway.
• ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY authenticates to that gateway.
• /status shows whether Claude Code is actually using the gateway URL and credential.
• A one-token POST /v1/messages request verifies the gateway before opening Claude Code.

LiteLLM’s Claude Code tutorial describes the non-Anthropic version of the same pattern:

• LiteLLM receives Claude Code requests in Anthropic Messages API format.
• It translates them to the target provider format.
• It forwards the request to OpenAI, Gemini, Azure OpenAI, Vertex AI, or another provider.
• It translates the response back to Anthropic Messages API format.
• With CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, Claude Code can populate /model from the gateway’s /v1/models endpoint.

Cherny added one caveat that matters for creators testing model swaps: the harness is tuned per model. In the same reply, he said the Claude Code team spends a lot of time dialing in model-specific tool design and prompting.

The clean distinction is credentials. Claude Code can point at a gateway, but Claude subscription credentials are not a general-purpose backend for third-party products.

Anthropic’s legal and compliance docs say OAuth authentication is for ordinary use of Claude Code and native Anthropic apps. The same page says developers building products with Claude capabilities should use API keys through Claude Console or a supported cloud provider, and that Anthropic does not permit third-party developers to offer Claude.ai login or route requests through Free, Pro, or Max credentials on behalf of users.

That makes bcherny’s denial in his false-positive reply narrower and more useful than the rumor cycle around it. He said Anthropic does not ban people for using other harnesses, while adding that accounts can still be banned for other reasons and that Anthropic is trying to drive down false positives.

Another classifier was already in the Claude Code news cycle. Anthropic said in its auto mode announcement that Pro, Max, and Team sessions will run auto mode by default starting August 14 unless users or admins pinned another default.

That classifier evaluates commands instead of asking users to approve every tool action. Anthropic’s production note says auto mode caught more dangerous actions than manual review in internal evaluations, and the launch post says Anthropic stopped charging Pro, Max, and Team users for the extra classifier tokens.

The community comparison landed immediately: steipete wrote in his OpenClaw reply that OpenClaw had shipped an auto classifier “for quite a while now.”

The account-classifier fix landed against a familiar pain point: locked accounts with little visible context.

kiaran_ritchie said in his ban report that he was banned with no reason given and no recourse. In his refund follow-up, he said Anthropic refunded the full month, calling the experience “weird and very frustrating.”

Anthropic’s safeguards appeals page says accounts can be banned for repeated usage-policy violations, unsupported-location account creation, or Terms of Service violations. The same page tells wrongly suspended users to log in with the banned account and fill out an appeal form, while noting longer-than-normal response times.