Anthropic Messages API ("Anthropic Skin")

Surplus speaks Anthropic's Messages API natively, so tools built for Anthropic — Claude Code and the Anthropic Agent SDK — can route through the marketplace with nothing but environment variables. No local proxy, no client-side translation.

This is the same idea as OpenRouter's "Anthropic skin": point the Anthropic client's base URL at Surplus, authenticate with a Surplus buyer key, and every request is routed to the cheapest healthy seller and settled on-chain like any other marketplace call.

It is model-agnostic. Routing is by the model string, not by family — so you can run Claude Code on Claude or GPT, GLM, Llama, Gemini, DeepSeek, or any other catalog model, all through the Anthropic wire.

Base URL

https://api.surplusintelligence.ai/anthropic

This is what you set as ANTHROPIC_BASE_URL. The Anthropic client appends /v1/messages (and the other paths below) to it. The skin lives under a dedicated /anthropic prefix so it never collides with the OpenAI-compatible surface at /v1/....

Authentication

Use a Surplus buyer API key (inf_…), sent either way the Anthropic ecosystem sends it:

Always set ANTHROPIC_API_KEY="" when using ANTHROPIC_AUTH_TOKEN, or the cached Anthropic credential can take precedence. An unauthenticated or non-buyer request returns an Anthropic-shaped 401 authentication_error (the skin does not issue x402/MPP 402 challenges — Claude Code can't satisfy them; use a buyer key with credits or an on-chain allowance).

Endpoints (under /anthropic)

Send a Message

curl https://api.surplusintelligence.ai/anthropic/v1/messages \
  -H "x-api-key: inf_xxx" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-4.8",
    "max_tokens": 1024,
    "system": "You are concise.",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

{
  "id": "msg_…",
  "type": "message",
  "role": "assistant",
  "model": "claude-opus-4.8",
  "content": [{"type": "text", "text": "Hello!"}],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {"input_tokens": 12, "output_tokens": 3}
}

Streaming

With "stream": true, the response is the standard Anthropic SSE event sequence:

event: message_start        → the Message skeleton
event: content_block_start  → per content block (text / thinking / tool_use)
event: content_block_delta  → text_delta / thinking_delta / signature_delta / input_json_delta
event: content_block_stop
event: message_delta        → stop_reason + cumulative usage
event: message_stop

Count Tokens

curl https://api.surplusintelligence.ai/anthropic/v1/messages/count_tokens \
  -H "x-api-key: inf_xxx" -H "Content-Type: application/json" \
  -d '{"model":"claude-opus-4.8","messages":[{"role":"user","content":"Hello"}]}'

{ "input_tokens": 14 }

List Models

{
  "data": [{"id": "claude-opus-4.8", "type": "model", "display_name": "Claude Opus 4.8", "created_at": "2025-01-01T00:00:00.000Z"}],
  "has_more": false,
  "first_id": "…",
  "last_id": "…"
}

Errors

Errors on /anthropic/* use the Anthropic envelope:

{ "type": "error", "error": { "type": "invalid_request_error", "message": "…" }, "request_id": "…" }

There is no Anthropic 402 — an out-of-credit buyer gets a 400 invalid_request_error (mirroring Anthropic's "credit balance too low" behavior).

Model names & aliases

Send any catalog model id. The skin additionally resolves:

• -latest tags: claude-opus-latest, claude-sonnet-latest, claude-haiku-latest, plus cross-family gpt-latest, glm-latest, deepseek-latest, gemini-latest, and more.
• Dated Anthropic IDs (e.g. claude-sonnet-4-5-20250929) → canonical.
• The OpenRouter ~anthropic/… floating-tag prefix.

A model must exist in the catalog and have a healthy offer to route; otherwise you get an Anthropic-shaped 404/overloaded_error. Call GET /anthropic/v1/models for the catalog and GET /api/markets for models with live seller liquidity.

Feature fidelity on non-Claude models

The wire is fully translated, but some Anthropic features are model-dependent:

• Works on every model: text, system prompt, multi-turn, streaming, tool use (tool_use/tool_result), images, stop reasons, usage/billing.
• Best-effort: extended thinking — non-Claude models don't emit Anthropic thinking blocks; the field is accepted and the request still works, but no reasoning blocks are returned. Thinking-signature continuity across multi-turn tool loops is preserved on providers that echo signatures and degrades gracefully elsewhere.

Routing controls

Everything the marketplace offers applies here too: provider pinning / allow-lists, provider_base_url, max_price_per_1m price caps, the /anthropic/min{N}/v1/messages minimum-discount mirror, and the buyer's trusted-only / allow-untrusted preference. See Minimum-Discount Routing and BYOK / Priority Provider.