Anthropic SDK migration

Stav serves the Anthropic Messages API at POST /v1/messages, plus POST /v1/messages/count_tokens.

Change two lines

import os
from anthropic import Anthropic

client = Anthropic(
    api_key=os.environ["STAV_API_KEY"],   # sk_stav_live_...
    base_url="https://api.stav.ai",       # no /v1 — the SDK adds it
)

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

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

A base URL ending in /v1 sends requests to /v1/v1/messages, which 404s.

Stav accepts the key as x-api-key (what the SDK sends) or Authorization: Bearer.

Model names

Use anthropic/claude-sonnet-4-5 or the bare claude-sonnet-4-5. Dated snapshots (claude-sonnet-4-5-20250929) and -latest names also resolve. auto, a Router (@acme/eu-only) and any other catalogue model work too.

Two paths, one endpoint

Claude models — proxied to Anthropic. The request goes to Anthropic's API as sent: thinking, cache_control, server tools, beta features and new fields all work. This applies when you name a Claude model and when auto or a Router selects one.

Every other model — translated to chat completions and dispatched like /v1/chat/completions.

Leaving out cache_control costs money, not correctness, so the call still succeeds. If your workload depends on the prompt cache, use a Claude model and watch for that header.

A Claude model is not sovereign. It runs at Anthropic. For EU-hosted inference, use a non-Claude model or a Router with a sovereignty floor. See Sovereignty.

Extended thinking

msg = client.messages.create(
    model="anthropic/claude-sonnet-4-5",
    max_tokens=8192,
    thinking={"type": "enabled", "budget_tokens": 4096},
    messages=[{"role": "user", "content": "Plan a Postgres 14 → 17 migration."}],
)

for block in msg.content:
    if block.type == "thinking":
        print("[reasoning]", block.thinking)

On Claude models Anthropic validates budget_tokens (at least 1024, below max_tokens) and Stav returns its error as is. See Reasoning.

Prompt caching

msg = client.messages.create(
    model="anthropic/claude-sonnet-4-5",
    max_tokens=1024,
    system=[{
        "type": "text",
        "text": LONG_SYSTEM_PROMPT,
        "cache_control": {"type": "ephemeral"},
    }],
    messages=messages,
)
print(msg.usage.cache_read_input_tokens)

Counting tokens

count = client.messages.count_tokens(
    model="anthropic/claude-sonnet-4-5",
    system=LONG_SYSTEM_PROMPT,
    tools=TOOLS,
    messages=messages,
)
print(count.input_tokens)

Not billed. For Claude models Anthropic does the count. For other models it is a local estimate from the model's tokenizer.

Scopes

A key with no scopes can call everything. A scoped key needs inference:messages for both endpoints.

Claude Code

export ANTHROPIC_BASE_URL="https://api.stav.ai"
export ANTHROPIC_AUTH_TOKEN="sk_stav_live_..."

Claude Code then sends every turn to /v1/messages on Stav and it is charged to your Stav wallet. With a Claude model everything works as before. With a non-Claude model, tools, streaming and thinking work; cache_control is left out as described above.

Migration checklist

1. Create a sk_stav_test_ key.
2. Set base_url to https://api.stav.ai — without /v1.
3. Run your tests against a Claude model. They should pass unchanged.
4. Try a non-Claude model or auto, and check X-Stav-Dropped-Params.
5. Set a sovereignty floor for the workspace.
6. Roll out with a sk_stav_live_ key.

Not available

• Files API, Message Batches API, Models API on the Anthropic path. Use GET /v1/models with an OpenAI-style client to list models.
• Server tools and computer use on non-Claude models.

Next steps

• Reasoning
• Tools & structured output
• Smart Router