Claude Code, Anthropic's official NPM package, is designed to send requests to the native /v1/messages endpoint, but many developers are unknowingly routing their requests through OpenAI's /v1/chat/completions format instead. This isn't a bug in Claude Code itself, but rather a configuration or environment issue that stems from one of six common mistakes in how developers set up their tools and dependencies.

Why Is Claude Code Using the Wrong API Endpoint?

When you install the official @anthropic-ai/claude-code package from NPM, it should automatically use Anthropic's Messages API protocol, which includes specific headers like anthropic-version and a native messages format for requests and responses. However, packet captures from some developers show requests hitting OpenAI's Chat Completions endpoint instead, which uses a completely different request structure and authentication method.

This protocol mismatch doesn't mean Claude Code is broken. Instead, it indicates that somewhere in the configuration chain, a conversion or package confusion has occurred. Understanding these six causes can help developers diagnose and fix the issue quickly.

What Are the Six Most Common Causes of This Configuration Error?

• Wrong Package Installation: Approximately 35% of cases involve accidentally installing a third-party OpenAI wrapper instead of the official @anthropic-ai/claude-code package. Similar-sounding packages like @gitlawb/openclaude, @aryanjsx/openclaude, cli-agent-openai-adapter, and claude-code-openai-wrapper all default to OpenAI's /v1/chat/completions endpoint.
• Routing Tool Interception: About 25% of cases involve tools like claude-code-router (CCR), a community tool designed to intercept Claude Code requests and forward them to various models. CCR starts a local proxy server and translates /v1/messages requests into /v1/chat/completions format before sending them upstream.
• OpenAI-Compatible Gateway Configuration: Roughly 20% of cases stem from pointing ANTHROPIC_BASE_URL to an LLM gateway like LiteLLM Proxy, OneAPI, or Bifrost that only exposes OpenAI-compatible endpoints. These gateways automatically rewrite requests to OpenAI format for upstream calls.
• Wrapper Environment Variables: About 10% of cases involve environment variables like CLAUDE_CODE_USE_OPENAI=1 or OPENAI_BASE_URL being set, which cause wrapper scripts to switch protocol modes even if the official package is installed.
• Shell Aliases or Wrapper Scripts: Approximately 7% of cases involve aliases or functions in ~/.bashrc or ~/.zshrc that intercept the claude command and redirect it to a wrapper tool.
• Packet Capture Misdeployment: In rare cases, about 3% of instances, the packet capture tool itself is deployed incorrectly, showing rewritten requests that don't reflect what Claude Code actually sent.

How to Diagnose and Fix Your Claude Code Configuration

• Verify the Official Package: Run which claude in your terminal to see the installation path, then run cat $(which claude) | head -20 to check if you're using @anthropic-ai/claude-code. If you see keywords like openclaude, claudex, or cli-agent-openai-adapter, you've found the root cause and should uninstall the wrapper and reinstall the official package.
• Check Your Environment Variables: Run env | grep -i ANTHROPIC to see what API configuration is active. Official Claude Code only recognizes ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL, and ANTHROPIC_MODEL. If you see OPENAI_BASE_URL, OPENAI_API_KEY, or CLAUDE_CODE_USE_OPENAI, you're likely using a wrapper.
• List Installed Claude-Related Packages: Run npm list -g --depth=0 | grep -i claude to see all globally installed packages with "claude" in the name. If npm list shows packages other than @anthropic-ai/claude-code, you may have accidentally installed a third-party wrapper.
• Search for Shell Aliases and Functions: Run alias | grep claude and type claude to check for aliases or functions that might be intercepting your Claude Code command. Also check ~/.bashrc, ~/.zshrc, and ~/.profile for any claude-related configurations.
• Verify Your Gateway Configuration: If you're using ANTHROPIC_BASE_URL to point to a custom gateway, verify that the gateway actually supports Anthropic's native /v1/messages endpoint. If it only supports /v1/chat/completions, you'll need to either switch gateways or accept the protocol conversion.

The correct configuration for official Claude Code should use environment variables like ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY, with an optional ANTHROPIC_BASE_URL pointing to https://api.anthropic.com (without the /v1 suffix). Requests should always include the anthropic-version header and use Anthropic's native messages format.

What Does This Mean for Developers Using Claude Code?

For most developers, this configuration issue is transparent. If your Claude Code setup is working correctly and returning the expected results, you're likely using the native /v1/messages endpoint without any issues. However, if you're experiencing unexpected behavior, slower response times, or compatibility issues, one of these six configuration problems may be the culprit.

The key takeaway is that Claude Code itself is not at fault. Instead, the issue lies in the surrounding environment, whether that's a third-party package, a routing tool, a gateway configuration, or shell-level interception. By systematically checking each of these six potential causes, developers can quickly identify and resolve the protocol mismatch and ensure they're using Anthropic's native API as intended.

Meanwhile, Anthropic continues to expand Claude's capabilities. The company is nearing a $900 billion valuation round, and Claude Security is now in public beta, offering enhanced protection features for enterprise users. These developments underscore Anthropic's commitment to making Claude more robust and enterprise-ready, even as developers work through configuration challenges with the underlying API integration.

" }