Use DeepSeek, Qwen, Kimi, GLM in Claude Code: config guide
Claude Code reads ANTHROPIC_BASE_URL (endpoint override), ANTHROPIC_AUTH_TOKEN (sent as Bearer) or ANTHROPIC_API_KEY (X-Api-Key), and ANTHROPIC_MODEL; they can also go in the env block of ~/.claude/settings.json. Live-tested: DeepSeek (Claude Code 2.1.302, text + tool calls pass). (source · verified)
Reference price via DeepSeek: DeepSeek V4.1 Flash: blended $0.523 per 1M tokens (official China site). (source · verified)
Claude Code reads ANTHROPIC_BASE_URL (endpoint override), ANTHROPIC_AUTH_TOKEN (sent as Bearer) or ANTHROPIC_API_KEY (X-Api-Key), and ANTHROPIC_MODEL; they can also go in the env block of ~/.claude/settings.json. Official docs
DeepSeek
shell
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_API_KEY=${YOUR_API_KEY}
export ANTHROPIC_MODEL=deepseek-flash Text call passes Tool call passes Tested 2026-10-17 with Claude Code 2.1.302 using the config above: text reply OK; Write created a file in a temp dir and Read returned the same content.
Alibaba Qwen (Model Studio)
~/.claude/settings.json
{
"env": {
"ANTHROPIC_AUTH_TOKEN": "YOUR_API_KEY",
"ANTHROPIC_BASE_URL": "https://dashscope.aliyuncs.com/apps/anthropic",
"ANTHROPIC_MODEL": "qwen3.7-plus"
}
} Pay-as-you-go uses a Bailian API key (sk-…); Coding Plan uses https://coding.dashscope.aliyuncs.com/apps/anthropic
Moonshot AI (Kimi)
No officially documented setup for this tool yet.
Kimi also offers an Anthropic-compatible Messages API (URL not on this page; unverified)
Zhipu GLM (Z.ai)
~/.claude/settings.json
{
"env": {
"CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.7",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2[1m]"
}
} Official route: run Coding Tool Helper; for GLM-5.2 1M context add the env above to ~/.claude/settings.json
Field names come from the linked official docs; replace placeholders with your key. Examples marked “to be verified” combine the tool’s documented fields with the provider’s OpenAI-compatible endpoint and were not run.