Claude Code

​

Routed mode

export ANTHROPIC_BASE_URL="https://your-relay/claude-code"
export ANTHROPIC_AUTH_TOKEN="sk-wr-your-relay-key"

​

Model discovery

​

Subagents

​

Settings

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://your-relay/claude-code",
    "ANTHROPIC_AUTH_TOKEN": "sk-wr-your-relay-key",
    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",
    "CLAUDE_CODE_GATEWAY_HINT_HEADERS": "1"
  }
}

​

Proxy mode

​

Prerequisites

1. Proxy mode enabled on your relay:
   curl -X PUT https://your-relay/api/settings/proxy-mode \
     -H "Authorization: Bearer $RELAY_ADMIN_TOKEN" \
     -H "Content-Type: application/json" \
     -d '{"enabled": true, "allowUnauthenticated": false}'
   
2. A relay key allowed to forward upstream credentials. Proxy requests are refused with 403 passthrough_forbidden unless the relay key has spec.passthroughAllowed: true. Set it when you create the key:
   curl -X POST https://your-relay/api/keys \
     -H "Authorization: Bearer $RELAY_ADMIN_TOKEN" \
     -H "Content-Type: application/json" \
     -d '{
       "metadata": {"displayName": "claude-code-proxy"},
       "spec": {
         "principal": {"kind": "user", "id": "<user-id>"},
         "policyId": "<policy-id>",
         "passthroughAllowed": true
       }
     }'
   
   or set it on an existing key with PUT /api/keys/by-id/{id}. Inference tokens never carry this flag, so proxy mode needs a relay key.
3. A policy that grants the Claude models Claude Code uses (Sonnet, Opus, Haiku) on that key. Without an explicit host header, Relay reads the model from the request and picks the upstream host from the key’s policy, so the models must be granted. The policy’s rate limits apply to proxied requests too.
4. (Optional) Payload logging turned on, if you want to capture the full request and response bodies. See Payload logging.

​

Setup

​

Option A: a wrapper command

#!/usr/bin/env bash
# Claude Code routed through Wyolet Relay.
export ANTHROPIC_BASE_URL="https://your-relay/claude-code"
export ANTHROPIC_CUSTOM_HEADERS=$'X-WR-API-Key: sk-wr-your-relay-key\nX-WR-Proxy-Mode: Proxy'
exec claude "$@"

​

Option B: Claude Code settings

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://your-relay/claude-code",
    "ANTHROPIC_CUSTOM_HEADERS": "X-WR-API-Key: sk-wr-your-relay-key\nX-WR-Proxy-Mode: Proxy"
  }
}

​

What you get

• Usage: every Claude Code request appears in GET /api/logs and the /api/usage/* endpoints of the control API, attributed to your relay key, with token counts, cost and timing.
• Payloads: with payload logging enabled, the full prompts and completions are captured and readable at GET /api/logs/{request_id}: what Claude Code sent and what the model returned.
• Attribution: the client’s own x-claude-code-session-id, x-claude-code-agent-id and x-claude-code-parent-agent-id headers, plus x-claude-code-request-class, -agent-type, -compaction and -context-compacted when CLAUDE_CODE_GATEWAY_HINT_HEADERS=1, are recorded on the usage event as extras, alongside client=claude-code. That is what lets you group and cost a whole session, or one subagent inside it.

​

Notes

• Setting ANTHROPIC_BASE_URL disables Claude Code’s MCP tool search by default; re-enable it with ENABLE_TOOL_SEARCH=true if you use it.
• Proxy mode does no translation: Claude Code’s Anthropic requests are forwarded byte for byte.
• When an upstream rejects a request on every key, Relay forwards that response as it stands (original status, Retry-After and body) instead of replacing it with a generic gateway error, so the client backs off on the provider’s own terms.

​

Known limits

• Silence before the first byte. Relay’s keepalive frames start once the upstream has sent its response headers. Some servers, Ollama among them, hold the headers until the first token is ready, so a model that spends longer than Claude Code’s idle limit (300 s by default) processing the prompt is cut off by the client before Relay can send anything. This happens with CPU-bound inference; on a GPU or a hosted upstream, prompt processing finishes in seconds. A smaller --tools set or --bare shrinks the prompt if a slow machine is all you have.
• Cross-shape prompt caching. When Relay translates to a non-Anthropic upstream, the system array is flattened into one instruction block, and cache_control markers have no equivalent there: OpenAI-compatible upstreams cache prefixes automatically and Gemini upstreams ignore them. Same-shape Anthropic upstreams receive both untouched. See Prompt caching.
• /context before the first turn. Relay answers count_tokens exactly only where the upstream has a counting endpoint (Anthropic); elsewhere the figure is calibrated from earlier turns of the session or estimated, and the X-WR-Token-Count response header says which.