Claude Code

Privatemode can be used as a model provider in Anthropic's Claude Code. This document explains how to set up Claude Code to use Privatemode.

Setting up the Privatemode proxy

Follow the API quickstart and choose Other languages to run the Privatemode proxy. Usually, it's advisable to enable prompt caching via --disable-prompt-cache=false. Check that the proxy is up with curl http://localhost:8080/v1/models.

Configuring Claude Code

Add the following configuration to your ~/.claude/settings.json to use Privatemode in every project. To try it in a single project first, put it into .claude/settings.local.json in that project instead.

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:8080",
    "ANTHROPIC_AUTH_TOKEN": "sk-privatemode",
    "ANTHROPIC_DEFAULT_MODEL": "glm-latest",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1048576",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "availableModels": [
    "glm-latest",
    "glm-flash-latest"
  ],
  "modelPicker": {
    "options": [
      {
        "model": "glm-latest",
        "label": "GLM"
      },
      {
        "model": "glm-flash-latest",
        "label": "GLM Flash"
      }
    ],
    "replaceBuiltInOptions": true
  }
}

The settings do the following:

• ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN point Claude Code at the proxy. The token is a placeholder, because the proxy holds your API key.
• ANTHROPIC_DEFAULT_MODEL sets the default model. availableModels restricts Claude Code to the models Privatemode serves: it uses Anthropic models for sub-agents by default, and with this list in place, sub-agents fall back to the session's active model instead. modelPicker replaces the built-in /model options with the Privatemode models.
• CLAUDE_CODE_MAX_CONTEXT_TOKENS sets the model's maximum context length of 1M tokens. Claude Code doesn't recognize the model and would otherwise assume a 200k context window.
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC disables optional telemetry and error reporting, and CLAUDE_CODE_ATTRIBUTION_HEADER omits Claude Code's attribution from the system prompt, which improves prompt-cache reuse.

Providing an API key per user

To use a separate API key per user, set ANTHROPIC_AUTH_TOKEN in ~/.claude/settings.json to your own key:

"ANTHROPIC_AUTH_TOKEN": "<your-api-key>"

The proxy must be started with --api-key-fallback=never, otherwise it uses the configured key instead.

Starting Claude Code with Privatemode

Claude Code reads its environment at startup, so open a new session after changing the settings:

claude

To use a different model for a single session, pass the --model flag, e.g.:

claude --model glm-flash-latest

You will be greeted with a screen like the following. You are now ready to "vibe code" confidentially 🥳

Reverting the setup

Remove the entries above from your settings file and stop the proxy. Claude Code then uses its default provider again.