Claude Code Setup (Anthropic Skin)

Point Claude Code (or the Anthropic Agent SDK) at Surplus and route every request through the marketplace — cheapest healthy seller — with no local proxy. Surplus implements Anthropic's Messages API surface (the "Anthropic skin"), so Claude Code talks to it the same way it talks to Anthropic.

It's also model-agnostic: you can run Claude Code on Claude or on GPT, GLM, Llama, Gemini, DeepSeek — any catalog model — through the same Anthropic wire.

For the full endpoint reference, see Anthropic Messages API.

1. Get a buyer key

Sign in on the Buy page with your wallet and create a buyer API key (inf_…). Fund it with deposited USDC credits or fiat credits. (Per-request x402 payment is not used by the skin — Claude Code authenticates with the buyer key.)

Claude Code uses two model slots

Main model

Your primary coding model. The one people remember to set.

Fast / small model

Background summarisation and cheap calls. Pin it too.

The fast/small slot handles background work and is easy to forget. Leaving it unpinned means part of your usage routes somewhere you did not choose.

2. Configure Claude Code

That's the whole required setup — three variables:

bash

export ANTHROPIC_BASE_URL="https://api.surplusintelligence.ai/anthropic"
export ANTHROPIC_AUTH_TOKEN="inf_your_key_here"
export ANTHROPIC_API_KEY=""        # important: must be empty so it doesn't override the auth token

Then run claude as usual. With no model variables set, Claude Code uses its built-in default models and the marketplace routes each request to the cheapest healthy seller for them.

Why ANTHROPIC_API_KEY=""? Claude Code prefers a cached ANTHROPIC_API_KEY (or a claude.ai login) over ANTHROPIC_AUTH_TOKEN. Blanking it forces your Surplus key to be used.

3. Choosing models (optional)

Setting models is optional. By default Claude Code uses its built-in models; the marketplace serves them like any other request. Set the variables below only when you want to pin a specific model, control cost, or run on a non-Claude model.

Claude Code has two model slots:

Variable	Slot	Used for
`ANTHROPIC_MODEL`	main	the primary agent loop
`ANTHROPIC_SMALL_FAST_MODEL`	small/fast	high-volume background tasks (titles, summaries, topic detection)

bash

export ANTHROPIC_MODEL="claude-opus-4.8"
export ANTHROPIC_SMALL_FAST_MODEL="gpt-4o-mini"

If you pin one slot, it's best to pin both (the small/fast slot is high-volume). Whatever model is used — a default or one you pin — must have a live offer in the marketplace; if it doesn't, those calls return a 503 / overloaded_error. Check the Market Prices page or GET /v1/models for what's live.

Newer Claude Code versions also read role-specific overrides (for the /model switcher and subagents), all optional:

bash

export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-4.8"
export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-4.6"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4.5"
export CLAUDE_CODE_SUBAGENT_MODEL="claude-opus-4.8"

You can also use floating tags — ~anthropic/claude-opus-latest, ~anthropic/claude-sonnet-latest, ~anthropic/claude-haiku-latest — which resolve to the current flagship.

Run Claude Code on non-Claude models

Because the skin routes by model name, just point the slots at any catalog model:

bash

export ANTHROPIC_MODEL="glm-4.7"            # or gpt-5.5, deepseek-v4-pro, gemini-3.1-pro, …
export ANTHROPIC_SMALL_FAST_MODEL="gpt-4o-mini"

A common cost-saver is a strong main model with a cheap small/fast model, e.g. ANTHROPIC_MODEL="claude-opus-4.8" + ANTHROPIC_SMALL_FAST_MODEL="gpt-4o-mini". Friendly -latest aliases work across families too (glm-latest, gpt-latest, deepseek-latest, …).

Text and tool use work on every model. Extended thinking is Claude-specific — non-Claude models won't return thinking blocks (the request still works). Smaller models follow the agent's tool protocol less reliably on complex multi-step tasks.

4. Verify

In Claude Code, run:

/status

It should show:

Auth token: ANTHROPIC_AUTH_TOKEN
Anthropic base URL: https://api.surplusintelligence.ai/anthropic

Ask it something simple ("reply with one word — pong"), then try a tool task ("read package.json and tell me the name field") to confirm the agent loop. You can watch usage on the Buy page as your credits settle.

5. Optional: routing controls

The skin honors the same marketplace controls as the OpenAI surface — set them in the request body (the Anthropic SDK passes through unknown fields), or use the minimum-discount mirror:

Provider pin / allow-list: "provider": "venice" or "provider": ["venice","openrouter"] (the OpenRouter {order|only} object form is accepted too).
Host pin: "provider_base_url": "https://api.openai.com/v1".
Minimum discount: send X-Min-Discount: 30, or set it once on the API key (min_discount_pct) so every request inherits it.
Minimum discount: set ANTHROPIC_BASE_URL="https://api.surplusintelligence.ai/anthropic/min50" to only route to sellers ≥50% cheaper than reference.

Troubleshooting

Symptom	Cause / fix
`401 authentication_error`	Missing/invalid key. Ensure `ANTHROPIC_AUTH_TOKEN` is your `inf_…` key and `ANTHROPIC_API_KEY=""`.
`503 overloaded_error` / "No available sellers for model 'X'"	Model `X` has no live offer right now. Pin that slot to a model with active liquidity (check `GET /v1/models` / Market Prices) — and if you pin one slot, pin both.
`404 not_found_error` (`minimum_discount_not_met`)	No seller meets your discount floor. The error names the best available discount — lower the floor, or wait for cheaper supply.
`400 invalid_request_error` mentioning credit	Buyer out of credit — top up with deposited USDC or fiat credits.
Claude Code ignores your key / uses claude.ai	`ANTHROPIC_API_KEY` not blanked, or a cached login takes precedence. Set `ANTHROPIC_API_KEY=""`.
High cost per turn	Claude Code's system prompt is large; the default models are big. Use a cheaper main model and a cheap small/fast model.

Anthropic Agent SDK

The same env vars work for the Anthropic Agent SDK / @anthropic-ai/sdk — set baseURL to https://api.surplusintelligence.ai/anthropic and the API key to your inf_… key. The SDK's x-api-key header is accepted.