Claude Code

Claude Code speaks the Anthropic Messages API. Point it at https://deference.si (no /v1) and give it your key as the auth token.

Set up

1. Create a key in API keys and export it as DEFERENCE_API_KEY.
2. Set the base URL, the token and model discovery in your shell profile.

export ANTHROPIC_BASE_URL="https://deference.si"
export ANTHROPIC_AUTH_TOKEN="$DEFERENCE_API_KEY"
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY="1"

$env:ANTHROPIC_BASE_URL = "https://deference.si"
$env:ANTHROPIC_AUTH_TOKEN = $env:DEFERENCE_API_KEY
$env:CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1"

3. Start claude. If you were signed in to Claude before, run /logout once so the saved login does not compete with the token.

To keep the setup out of your shell profile, put it in the env block of your user file, ~/.claude/settings.json (on Windows, %USERPROFILE%\.claude\settings.json). The block does not expand variables, so the key goes in as text.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://deference.si",
    "ANTHROPIC_AUTH_TOKEN": "sk-df-...",
    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1"
  }
}

Verify

1. Check the route and the key with one request. A reply that starts with {"id":"msg_ means it works.

curl https://deference.si/v1/messages \
  -H "Authorization: Bearer $DEFERENCE_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-5.5",
    "max_tokens": 1,
    "messages": [{ "role": "user", "content": "." }]
  }'

2. Start claude, send a short prompt, then run /status. It lists the Anthropic base URL as https://deference.si and names the auth token. A login method that names claude.ai means Claude Code did not pick up the token.
3. Run claude --debug to see the model discovery lines in the debug log.
4. Open Activity. The request is the top row, and its client reads Claude Code.

Pick a model

Claude Code asks for a model class (Opus, Sonnet or Haiku) and resolves it to Anthropic's own model id. Pin each class to a Deference id so the request names a model in the catalog.

• A model named in a subagent's definition or invocation wins over CLAUDE_CODE_SUBAGENT_MODEL. To pin every subagent to one id, also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 on Claude Code 2.1.257 or later.
• With discovery on, /model lists the catalog ids that contain claude or anthropic. Select any other model with /model <id> or claude --model <id>, using its full id.
• A model Claude Code does not recognize is assumed to have a 200K context window. Append [1m] to the id, as in anthropic/claude-sonnet-5.5[1m], to ask for 1M. Deference strips that suffix, looks the model up without it, and sends the catalog id. The provider does not see [1m].
• Claude Code is built for Claude models. Other models in the catalog answer, but tool use and thinking can behave differently.
• On a team, set allowedProviders to ["customEndpoint"] in Claude Code's managed settings on v2.1.285 or later. The same file's env block must set ANTHROPIC_BASE_URL to https://deference.si; Claude Code then refuses a session pointed anywhere else.

Use it in other Claude clients

The Claude Agent SDK and the Claude Code GitHub Action read the same variables. The VS Code extension reads them from its own setting, claudeCode.environmentVariables, not from settings.json. The Claude desktop app ignores ANTHROPIC_BASE_URL.

Connect the MCP server

Connect the MCP server and Claude Code checks your credit, runs models within a budget you set, and creates, rotates and revokes its own capped keys. You sign in once.

Troubleshooting

Messages errors carry a type and a message, not a code, so each row names the status, the type and how the message starts.