Use OpenClaw with open-source models hosted in the EU

OpenClaw is a self-hosted AI assistant. A gateway process runs on your laptop, a home server or a VPS, and you reach your agents through messaging channels and a web chat. The agents can browse, run tools, manage files and work through scheduled tasks on your behalf.

Because OpenClaw runs on your own infrastructure, you choose the model provider. Alongside its bundled providers, it supports custom providers and proxies that speak the OpenAI Chat Completions or Anthropic Messages format. That is how you connect it to inference.eu.

OpenClaw reads its settings from ~/.openclaw/openclaw.json, written in JSON5. Under models.providers, the inference-eu entry sets baseUrl to the inference.eu base URL ending in /v1, api to openai-completions and apiKey to ${INFERENCE_EU_API_KEY}, which OpenClaw resolves from the environment. mode: "merge" keeps the bundled providers available next to yours.

Each model needs at least an id matching inference.eu. Agents refer to models as provider/model, so agents.defaults.model.primary becomes inference-eu/kimi-k3. Onboarding can write the same settings with --custom-base-url and --custom-model-id. For the Anthropic format, use anthropic-messages and the base URL without /v1.

OpenClaw agents plan over many steps, call tools and keep long conversations going across channels, so they need models with dependable tool calling and good long-context handling. kimi-k3 is a strong primary model for general assistant work, and glm-5.3 is a good alternative with careful multi-step reasoning.

For agents that mostly write or review code, qwen3-coder-480b-a35b-instruct is a good fit. gpt-oss-120b answers quickly and suits lightweight agents, triage and short replies in chat. If a model accepts images, add input: ["text", "image"] to its entry so attachments reach it as images.

A 401 error means the gateway process cannot see the key. Services started by systemd or launchd do not read your shell profile, so set INFERENCE_EU_API_KEY in the service environment or under env in the config. A 404 usually comes from a baseUrl with an extra path.

"Unknown model" means primary points at an id missing from models.providers.inference-eu.models; add the entry, since agent aliases alone do not register a model. If an agent stops calling tools after a model switch, check the model supports tool calling and run openclaw doctor to catch stale settings.

With inference.eu as the model provider, the messages, files and tool output your agents send are processed in EU data centres and never used for training. We sign a data processing agreement under Article 28 GDPR, publish our subprocessor list and bill in euros.

OpenClaw also talks to the messaging channels you connect and to any tools or websites your agents use; those services handle data under their own terms. Remove bundled providers you do not need from your agents' fallbacks, and host the gateway in the EU, so that model traffic and stored sessions stay in Europe.