Anthropic SDK migration

Stav speaks the Anthropic Messages contract as a first-class API surface, in parallel with the OpenAI Chat Completions one. If you already use the official Anthropic SDK, the migration is two lines.

The change

Python

import anthropic

client = anthropic.Anthropic(
    base_url="https://api.stav.ai/v1",       # ← was api.anthropic.com
    api_key=os.environ["STAV_API_KEY"],      # ← was ANTHROPIC_API_KEY
)

Node / TypeScript

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://api.stav.ai/v1",         // ← was api.anthropic.com
  apiKey: process.env.STAV_API_KEY,
});

Every other call — messages.create, messages.stream, tool use, extended thinking, prompt caching, vision — works as published in the Anthropic SDK.

How Stav serves Anthropic requests

There are two paths under the hood, and which one you get depends on the model identifier:

Both paths support streaming, tools, and structured output. The decision is per-request, based on the model you pass.

Model identifiers

client.messages.create(model="claude-sonnet-4-5", ...)              # → Anthropic upstream
client.messages.create(model="claude-haiku-4-5-20251001", ...)      # → Anthropic upstream
client.messages.create(model="anthropic/claude-sonnet-4-5", ...)    # alias, same as above

client.messages.create(model="qwen3-32b", ...)                      # → sovereign worker
client.messages.create(model="auto", ...)                           # → Smart Router
client.messages.create(model="meta-llama/llama-4-maverick", ...)    # → sovereign worker

The catalogue at /v1/public/models lists every identifier with its sovereignty tier and pricing.

What's identical (for claude-* models)

What's identical (for sovereign models)

The translation layer covers the common surface area, but with two natural limits:

• Anthropic-only features map to their closest equivalent. thinking becomes equivalent reasoning steering on models that support it (most large sovereign models do). cache_control markers are honoured by Stav's prompt-caching layer.
• Beta flags that require Anthropic-specific implementations (e.g. computer-use tools, which require Anthropic's vision-action models) only work with claude-* models.

If you need a feature that's specifically Anthropic's implementation, target a Claude model. If you need any compatible model on the same SDK and call shape, point at auto or any sovereign model identifier — the translation handles the rest.

What's different — Stav extensions

The same set of extensions as the OpenAI surface. You can use them with the Anthropic SDK by passing them as extra_headers or in the body (the Anthropic SDK doesn't have a built-in extra_body parameter — see below).

Sovereignty tier on every response

response = client.messages.with_raw_response.create(
    model="claude-sonnet-4-5", max_tokens=1024,
    messages=[{"role": "user", "content": "..."}],
)
tier = response.headers["X-Stav-Sovereignty-Tier"]   # "routed-commercial" for claude-*
parsed = response.parse()

model="auto" — Smart Router

Same semantics as the OpenAI side. model="auto" lets the Smart Router pick the best model per request based on team policy, including the option to land on a Claude model. The response carries X-Stav-Selected-Model so you know what ran.

:fast / :precision tier suffix

Append to any sovereign model identifier. Claude models do not have a tier suffix — Anthropic chooses the quantization for their hosted endpoints.

Sovereign built-in tools

Pass stav.*-named tools in the tools array; Stav executes them server-side and folds the result back into the conversation in the same request. See Built-in tools.

client.messages.create(
    model="auto",
    max_tokens=1024,
    messages=[{"role": "user", "content": "What did the EU AI Act say about general-purpose model obligations?"}],
    tools=[{"name": "stav.web_search", "description": "Stav sovereign web search", "input_schema": {"type": "object", "properties": {}}}],
)

(The Anthropic schema requires an input_schema; Stav ignores its content for stav.* tools — the canonical schema is server-side.)

Stav-specific request parameters

When using the Anthropic SDK, Stav extensions go in extra_headers:

client.messages.create(
    model="auto",
    max_tokens=1024,
    messages=[...],
    extra_headers={
        "X-Stav-Sovereignty-Required": "true",
        "X-Stav-Max-Tool-Iterations": "3",
    },
)

Migrating a real codebase — checklist

Known compatibility gaps

• Computer-use tools. Only work with claude-* models. Stav doesn't yet have a sovereign equivalent (it's on the sovereign-tools roadmap; see the sovereign-tools strategy doc).
• PDF / files upload via Anthropic's Files API. Use the stav.file_search ingestion pipeline (/v1/vector_stores) instead.

Next steps

• OpenAI SDK migration — if you also have OpenAI-SDK call sites.
• Built-in tools — sovereign tools accessible through both SDKs.
• Smart Router — the cost-down recommendation that follows the base-URL swap.