Claude Code on AI models run in Europe

Point Claude Code at Evanesca's Anthropic-format endpoint and run the terminal agent on open models run in Europe, tool calls included. The env block of ~/.claude/settings.json: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and one model variable per slot.

Tool documentation read on 12 Oct 2026

SoonOpening soon: what follows describes the service as it will work.

Tool: Claude Code · Coding agent
API format: Anthropic Messages
Base URL: https://evanesca.app/api
Your prompts: From your machine or server, straight to Evanesca

Setup

1. Give Claude Code the endpoint and a model for every slot

Claude Code adds /v1/messages to ANTHROPIC_BASE_URL, so the URL stops at /api. Fill every model slot, the background one (Haiku) included: a slot left empty asks for a Claude model, which Evanesca does not serve, and that call fails. Here DeepSeek V4.1 Flash drives the work and Gemma 4 26B handles the background tasks. Keep the key in your user settings, never in a project file under version control.

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://evanesca.app/api",
    "ANTHROPIC_AUTH_TOKEN": "<your Evanesca key>",
    "ANTHROPIC_MODEL": "deepseek-v4.1-flash",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-v4.1-flash",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4.1-flash",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gemma-4-26b",
    "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-v4.1-flash",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

2. Check the settings, then start

In a session, /status prints the base URL and the token in use, and /model moves between the slots you filled. To try it once without touching your settings, export the same variables in a shell.

Shell

export ANTHROPIC_BASE_URL=https://evanesca.app/api
export ANTHROPIC_AUTH_TOKEN=$EVANESCA_API_KEY
export ANTHROPIC_DEFAULT_HAIKU_MODEL=gemma-4-26b
claude --model deepseek-v4.1-flash

What the Anthropic-format endpoint does, field by field: POST /messages.

Where these settings come from

These settings come from Claude Code's official documentation, read on 12 Oct 2026. This guide has not been run against Evanesca yet: it will be checked when the API opens.

Which model to pick

Evanesca's models from which a real tool call came back at their operator. This guide uses DeepSeek V4.1 Flash and Gemma 4 26B, picked from the catalogue for their properties. Prices per million tokens, input / output, margin and VAT included.

Model ID	Operator and place	Context	Reasoning	Price
`mistral-large-4`	EU region of Mistral AI	512k	—	€2.5133 / €7.7246
`deepseek-v4.1-flash`	Germany (AKI.IO)	1M	Yes	€1.008 / €3.024
`glm-5.3`	Germany (AKI.IO)	512k	Yes	€3.024 / €9.24
`qwen3.8-27b`	Scaleway — Paris, France	256k	Yes	€1.008 / €5.544
`deepseek-v4-flash`	Scaleway — Paris, France	256k	Yes	€0.672 / €1.344
`apertus-v1.5-70b`	Switzerland (Infomaniak)	100k	—	€1.2936 / €4.62
`glm-5.2`	Scaleway — Paris, France	256k	Yes	€3.024 / €9.24
`qwen3.6-35b`	Scaleway — Paris, France	256k	Yes	€0.42 / €2.52
`gemma-4-26b`	Germany (AKI.IO)	256k	—	€0.168 / €0.84
`mistral-small-4-119b`	EU region of Mistral AI	256k	—	€0.2772 / €1.1088
`qwen3.5-397b`	Scaleway — Paris, France	250k	Yes	€1.008 / €6.048
`gpt-oss-120b`	Scaleway — Paris, France	—	Yes	€0.252 / €1.008
`qwen3-coder-30b`	Scaleway — Paris, France	256k	—	€0.336 / €1.344
`qwen3-235b`	Scaleway — Paris, France	250k	—	€1.26 / €3.78
`mistral-small-3.2`	Scaleway — Paris, France	—	—	€0.252 / €0.588
`llama-3.3-70b`	Scaleway — Paris, France	—	—	€1.512 / €1.512
`pixtral-12b`	Scaleway — Paris, France	128k	—	€0.336 / €0.336

Reasoning tokens are billed as output. Where each model runs and what its operator keeps: Models and Europe.

Good to know

Anthropic's documentation states that it does not support Claude Code with non-Claude models through a gateway. This setup relies on Evanesca translating the Messages format; Anthropic does not stand behind it.
Every file read and every command is a tool call: keep to models whose tool calling is measured (table below).
Set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the window of the model you chose: for a model ID it does not know, Claude Code assumes 200,000 tokens.
Features that run on Anthropic's own servers stay out of reach through a gateway: the WebSearch tool, Remote Control, claude.ai connectors and voice dictation.
A model's reasoning does not come back as thinking blocks, which would need a signature only Anthropic can give; while it thinks, ping events keep the stream open. Cache markers (cache_control) are accepted and ignored.

Where your prompts go

On Evanesca's side, the model runs at a European operator, named with its place in the x-evanesca-operator and x-evanesca-location headers of every response, and Evanesca keeps no copy of prompts or answers — only the billing data of each call. What each operator keeps is set out in Models and Europe; the host of Evanesca's servers is named in the legal notice.

Sources

Official documentation, read on 12 Oct 2026:

Frequently asked questions

Does Claude Code work with Evanesca?

Yes, through its base URL setting: The env block of ~/.claude/settings.json: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and one model variable per slot. These settings come from Claude Code's official documentation, read on 12 Oct 2026. This guide has not been run against Evanesca yet: it will be checked when the API opens.

Which Evanesca model should I use with Claude Code?

One whose tool calling is measured. This guide uses DeepSeek V4.1 Flash and Gemma 4 26B; the 17 candidates are in this page's table, with their context window and price.

Do my prompts stay in Europe with Claude Code?

Model requests leave your machine for ANTHROPIC_BASE_URL. Claude Code still sends telemetry and update checks to Anthropic unless CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is 1 (set in the sample), and its WebFetch tool checks each domain with api.anthropic.com unless skipWebFetchPreflight is on. On Evanesca's side, the model runs at a European operator, named with its place in the x-evanesca-operator and x-evanesca-location headers of every response, and Evanesca keeps no copy of prompts or answers — only the billing data of each call. What each operator keeps is set out in Models and Europe; the host of Evanesca's servers is named in the legal notice.

One key, and the tool runs on models in Europe.

Prepaid credit, no subscription: the same key works with every guide.

Get an API key All integrations