Run Claude Code on Stav

Claude Code is Anthropic's coding agent for the terminal. It reads your code, edits files and runs commands. It can send its model traffic to any gateway that speaks the Anthropic Messages API, so it runs on Stav with configuration only.

Point it at Stav and every turn is billed to your Stav credits, shows up in your request log, and follows your workspace policy. Pick a model that is not a Claude model and the work runs where that model's Inference Provider runs — for example in France.

Tested with Claude Code 2.1.296 on macOS.

Before you start

• A Stav API key. See Authentication.
• Claude Code installed. See the Claude Code quickstart.

Set it up

Add an env block to ~/.claude/settings.json. It applies to every project and to Claude Code's background agents:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.stav.ai",
    "ANTHROPIC_AUTH_TOKEN": "sk_stav_live_...",
    "ANTHROPIC_MODEL": "zai-org/glm-5-2"
  }
}

Or, for one terminal session:

export ANTHROPIC_BASE_URL="https://api.stav.ai"
export ANTHROPIC_AUTH_TOKEN="sk_stav_live_..."
export ANTHROPIC_MODEL="zai-org/glm-5-2"

• Write the base URL without /v1. Claude Code adds /v1/messages itself.
• Use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. It sends your key as Authorization: Bearer and takes effect at once. (ANTHROPIC_API_KEY also works, but Claude Code asks you to approve it first.)
• Never put the key in a project's .claude/settings.json. That file is committed to your repository.

Start claude and run /status. The Anthropic base URL line should show https://api.stav.ai, and Auth token should name ANTHROPIC_AUTH_TOKEN.

Stav credits pay for it. While the key is set, Claude Code does not use your Claude subscription. Every turn is charged to your Stav credits.

Pick a model

Set ANTHROPIC_MODEL to any chat model in the Stav catalogue, by its full slug.

A named model can be served by more than one Inference Provider, and they are not all in the EU. To keep every request in Europe, set a sovereignty floor or a compute region for your workspace. See Keep traffic sovereign.

Model auto sends the request to your workspace's default router, which the Smart Router drives. It has nothing to do with Claude Code's own "auto mode", which is a permission setting.

What works

Tested on zai-org/glm-5-2:

A Claude model gets the request exactly as Claude Code sent it, so Claude Code's features behave as they do on Anthropic's own API. We tested one streamed turn on claude-haiku-5-5.

Tips

• Output limits are handled for you. Claude Code asks for up to 32,000 output tokens per turn. Where an Inference Provider allows fewer, Stav sends the provider's own limit and names it in X-Stav-Clamped-Params (for example max_tokens=16384 on GLM-5.2). To choose a lower ceiling yourself, set CLAUDE_CODE_MAX_OUTPUT_TOKENS.
• Tell Claude Code the real context window. Claude Code does not know non-Claude models and assumes 200,000 tokens. For a model with a larger window, set CLAUDE_CODE_MAX_CONTEXT_TOKENS, for example "1000000" for GLM-5.2. The model page in the catalogue shows the window.
• The warning about an unknown model is harmless. Claude Code prints that the model "isn't described by this version's model catalog". It is Claude Code's own list, not an API error. The run continues.
• Model discovery only finds Claude models. With CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, Claude Code reads Stav's /v1/models, but keeps only names that contain claude or anthropic. Set other models with ANTHROPIC_MODEL or /model <slug>.
• Some traffic does not go through Stav. Claude Code still sends telemetry and update checks to Anthropic and others. To turn that off, add "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1" to the env block. This also turns off auto-updates.
• The VS Code extension reads its own settings. Put the same variables in claudeCode.environmentVariables in VS Code's user settings. See Configure each surface.

Troubleshooting

For Stav's error codes, see Errors.

Check it worked

Open Your workspace → Monitor → Request log. Each turn is one row, with the model, Inference Provider, country and cost. A short task is usually 3–5 rows. See Request log and prompt capture.