Integration guide · Coding agent

Claude Code on EU-hosted open models

Point Claude Code at Frontière AI's Anthropic-compatible endpoint and run it on EU-hosted open models, tool calls included. The env block of ~/.claude/settings.json: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and the ANTHROPIC_DEFAULT_*_MODEL variables.

Tested on 13 October 2026. Claude Code 2.1.296 against our /v1/messages endpoint: streamed answers, seven tool-calling turns (Read, Bash) and a correct final answer, with gpt-oss-120b on OVHcloud.

API format

Anthropic Messages

Base URL

https://getfrontiereai.eu/api

Where it goes

The env block of ~/.claude/settings.json: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and the ANTHROPIC_DEFAULT_*_MODEL variables.

Your prompts

Straight from your machine to us

Setup

// ~/.claude/settings.json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://getfrontiereai.eu/api",
    "ANTHROPIC_AUTH_TOKEN": "sk-front-...",
    "ANTHROPIC_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3.6-35b",
    "CLAUDE_CODE_SUBAGENT_MODEL": "glm-5.2",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "256000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

export ANTHROPIC_BASE_URL=https://getfrontiereai.eu/api
export ANTHROPIC_AUTH_TOKEN=sk-front-...
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.6-35b
claude --model glm-5.2

Which model to pick

24 live models have tool calling measured by a real call through the gateway. Models labeled EU sovereign run at European providers; fast-access models are served through US-controlled infrastructure. Prices are per 100,000 tokens, margin and VAT included.

Good to know

• Anthropic's own documentation says it does not support routing Claude Code to non-Claude models through any gateway. It works because Frontière AI translates the Messages format faithfully; it is not an Anthropic-supported setup.
• Pick models with measured tool calling (table below): Claude Code is an agent, every file read and command is a tool call.
• Set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the model's real window: for an unknown model id Claude Code assumes 200,000 tokens.
• Features that rely on Anthropic's servers do not work through a gateway: the WebSearch tool (an Anthropic server-side tool), Remote Control, claude.ai connectors and voice dictation.
• Reasoning traces are not shown as thinking blocks: they have no valid Anthropic signature. Long reasoning keeps the stream alive with ping events.
• Prompt-cache markers (cache_control) are accepted and ignored; the providers cache prompt prefixes on their own, and cached tokens are reported in usage.

Where your prompts go

Model requests go from your machine straight to ANTHROPIC_BASE_URL. Claude Code still sends telemetry and update checks to Anthropic unless CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 (in the snippet), and its WebFetch tool checks each hostname with api.anthropic.com unless skipWebFetchPreflight is set.

On our side, a model labeled EU sovereign runs at a European provider, and the gateway itself runs in Paris on Vercel, a US company — our privacy page says so. Billing keeps the model, the token counts and the amount, never the prompt or the reply.

Sovereignty →Privacy →

Sources

Official documentation, checked on 12 October 2026:

• Claude Code — Connect to an LLM gateway
• Claude Code — Gateway compatibility guide
• Claude Code — Settings and environment variables

FAQ

Other guides

• OpenCodeCoding agent
• ClineCoding agent
• Kilo CodeCoding agent
• ContinueCoding assistant