Claude Code — custom API endpoint setup

Point Claude Code at a third-party Anthropic-compatible endpoint via settings.json.

The one thing people get wrong

Claude Code speaks the Anthropic Messages protocol. The base URL is the site root with no /v1:

CORRECT:  https://apimaster.ai
WRONG:    https://apimaster.ai/v1

With /v1 appended, Claude Code builds https://host/v1/v1/messages, gets a 404, and falls back to prompting you for an Anthropic login — which looks like an auth bug but isn't.

Configure

Windows — Win + R → %userprofile%\.claude macOS / Linux — ~/.claude

Create or edit settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://apimaster.ai",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_API_KEY",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
    "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1"
  }
}

CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS matters: third-party endpoints often don't implement Anthropic's beta headers, and leaving them on causes requests to fail in ways the error message doesn't explain.

Run

claude
/model claude-sonnet-4-6

Troubleshooting

Symptom	Cause	Fix
Prompts for Anthropic login	/v1 in the base URL	Use the site root
401	Variable name mismatch	Swap ANTHROPIC_AUTH_TOKEN ↔ ANTHROPIC_API_KEY — which one is read depends on version
Unknown model	Id not served here	Check your endpoint's model list
Works, then breaks after update	New betas re-enabled	Confirm the two CLAUDE_CODE_* flags survived

If your endpoint only speaks OpenAI protocol

Then Claude Code can't reach it directly. Use claude-code-router as a translation layer, and give it the full endpoint path so it doesn't mis-join:

{
  "Providers": [{
    "name": "custom",
    "api_base_url": "https://apimaster.ai/v1/chat/completions",
    "api_key": "YOUR_API_KEY",
    "models": ["claude-sonnet-4-6"]
  }]
}

Then set ANTHROPIC_BASE_URL to your local CCR address.

Check whether you need this at all — run check_endpoint.py below. If the Anthropic Messages test passes, connect directly and skip CCR entirely.

---

Verify before you debug

Before blaming the agent, confirm the endpoint itself works. This repo ships a zero-dependency checker that tests both protocols and prints the exact base_url to paste into each tool:

python check_endpoint.py --url https://apimaster.ai --key YOUR_API_KEY

It reports which protocol the endpoint speaks, TTFT, the advertised model list, and whether your model id is actually on it.

---

Other agents

Configs for Claude Code, Kilo Code, OpenCode, Aider, Codex, Cursor, LiteLLM and LangChain: ai-coding-agent-api-configs

---

Endpoint used in the examples

Examples use https://apimaster.ai/v1, an OpenAI-compatible endpoint fronting multiple model providers with one key. Every config here works with any OpenAI-compatible endpoint — substitute your own, including a self-hosted new-api or LiteLLM proxy.

Get a key · Docs

License

MIT