Skip to content

Use with Claude Code

Atlas implements the Anthropic Messages API as a first-class surface, so an app using the Anthropic SDK, the Claude Agent SDK, or Claude Code works against Atlas by changing ANTHROPIC_BASE_URL (plus key and model name) and nothing else.

Terminal window
ANTHROPIC_BASE_URL=http://localhost:8080 ANTHROPIC_API_KEY=<your key> claude

The key is the one Atlas printed on first start (or any you mint with atlas keys create). The same two variables work for the Anthropic Python/TypeScript/Go SDKs and the Claude Agent SDK.

Operators define aliases that map a requested model to a deployed one, so SDK and tool defaults resolve. For example, an operator might map claude-sonnet-4-6 → qwen3-coder-72b. Two ways to control which local model Claude Code uses:

  • Aliases — request the Claude model name your tooling defaults to; the operator’s alias resolves it to a local model.
  • Default-model env vars — set ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, or ANTHROPIC_DEFAULT_HAIKU_MODEL to a model Atlas serves.
  • Streaming, multi-turn, system prompts, and tool use (the agent tool loop) — the full Anthropic SSE event sequence, including tool-call streaming that agent SDKs depend on.
  • Thinking blocks for reasoning-capable models (DeepSeek-R1, Qwen3, QwQ, GLM, gpt-oss, …), mapped to Anthropic thinking content blocks. thinking.type: adaptive (Claude Code’s default) is accepted. Non-reasoning models simply return no thinking blocks.
  • count_tokens for budgeting, and GET /v1/models listing deployed models and aliases.

See API compatibility for the precise surface and what is intentionally out of scope.