Claude Code and Anthropic SDKs

The gateway serves the Messages API itself: each request is translated to a chat request, runs through the same admission, budgets, boons and billing as /v1/chat/completions, and the reply is translated back. Your registered models answer; no Anthropic service is involved.

Set up Claude Code

1. 1

   Point it at the gateway

   Set ANTHROPIC_BASE_URL to the data-plane origin without a path (the client adds /v1/messages), and a gateway key as the credential:

   export ANTHROPIC_BASE_URL=http://localhost:8088
   export ANTHROPIC_API_KEY=sk_...
   claude
   

   ANTHROPIC_API_KEY is sent as x-api-key and ANTHROPIC_AUTH_TOKEN as Authorization: Bearer; the gateway accepts both, and an identity token works in either. To keep the settings, put them in the env block of ~/.claude/settings.json, not in a project's .claude/settings.json, which is usually committed.

2. 2

   Pick the model that answers

   Claude Code sends Claude model ids that your gateway does not serve. Map them with one of the options in the next section.

3. 3

   Check it

   curl -s "$ANTHROPIC_BASE_URL/v1/messages" \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2023-06-01" \
     -H "content-type: application/json" \
     -d '{"model": "qwen3-coder", "max_tokens": 64,
          "messages": [{"role": "user", "content": "ping"}]}'
   

   A message object back means the key, the model and its backend all work. A not_found_error means the model name did not resolve.

The Anthropic SDKs take the same two values:

from anthropic import Anthropic

client = Anthropic(base_url="http://localhost:8088", api_key="sk_...")
message = client.messages.create(
    model="qwen3-coder",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)
print(message.content[0].text)

Map Claude model ids to your models

Use any of these; they combine.

• Pin the model on the client. Set ANTHROPIC_MODEL (or claude --model) for the main session, and ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL and CLAUDE_CODE_SUBAGENT_MODEL for the rest. The value is any gateway model name, alias or auto.
• Add Claude ids as aliases. Add the ids the client sends, such as claude-sonnet-4-6, to the aliases of the model that should answer. Aliases match exactly, so a client release with new ids needs new aliases.
• Set a default for Anthropic clients. In Settings › Routing, choose a model under Default for Anthropic clients, and save. Any name that does not resolve on /v1/messages is then served by that model; it can be auto. Over the API it is messages_default_model on PUT /api/v1/settings/auto-router, and an empty string clears it.

For each request the gateway uses auto as sent, then a name or alias it knows, then the default. With none of these it returns 404 not_found_error with model: <name>. Routing, pools, budgets and usage follow the model that served the request, while the response's model field echoes the name the client sent. The default does not stand in for a model that exists but is disabled; that returns 403.

What is translated

Replies come back as a message object with thinking (from the model's reasoning output), text and tool_use blocks, and stop_reason of end_turn, max_tokens or tool_use. Streams send ping every 15 seconds while the backend is quiet.

Usage follows Anthropic's fields. input_tokens is the input the backend did not serve from its prefix cache, cache_read_input_tokens the part it did, and cache_creation_input_tokens is always 0, since a prefix cache has no write charge. A stream's message_start is sent before usage is known, so its counts are 0; the final message_delta carries them.

Token counting

POST /v1/messages/count_tokens takes the same body (max_tokens optional) and returns {"input_tokens": 1834}, the gateway tokenizer's estimate of the message text. It takes no slot, reserves no budget and writes no usage row.

Not supported

• Prompt cache control. cache_control markers are dropped. Caching is up to the backend's own prefix cache, and usage reports what it served.
• Extended thinking budgets. thinking and output_config.effort are read as a reasoning effort level (disabled is none; a budget is bucketed into low, medium or high) and reach the model only through its effort map or a model family. The budget itself is not passed on. Reasoning a model produces still comes back as thinking blocks.
• Images inside tool results. Only the text of a tool_result reaches the model.
• Server-side tools. Web search, code execution, computer use, container, mcp_servers and the batch API. The model's own boons, including the gateway tool loop, still run.
• Which stop sequence fired. stop_sequence is always null.
• A per-tenant default. The default for Anthropic clients is one gateway-wide setting.