Skip to content

Run AI agents on European infrastructure

An agent needs a model, a machine, tools and a region. Munito AI runs all four: open models behind the OpenAI and Anthropic APIs, and a cloud computer for each agent.

Updated Read as Markdown

An agent needs four things, and we run all four

A model to think with. A machine to work on. Tools it can call. A region the work stays in. Most platforms sell you the first one and leave the rest to you.

Munito AI runs the inference and the sandboxes on the same account, behind one key, in the EU, with zero data retention by default.

The parts, and where each one lives

The agent needsMunito AI gives it
A modelOpen models behind the OpenAI and the Anthropic APIs
A machineA sandbox: a cloud computer started in about a second
ToolsTool calling, structured outputs, and an MCP server for your account
A long memoryA 1M-token context on DeepSeek-V4-Flash, DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash
A place to runThe EU, the US, your own cloud, or on-premises

Keep the harness you already have

The API answers in two shapes, on the same key. An OpenAI-shaped agent changes a base URL. An Anthropic-shaped agent, Claude Code included, changes a base URL. Neither needs a proxy.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.eu.munito.ai/inference/v1",
    api_key="munitoai_...",
)

Eight harnesses have a guide of their own: Claude Code, Codex, OpenCode, Hermes Agent, pi, OpenClaw, DeepSeek Harness and Prime Agent. See the agent guides and Compatibility.

Give the agent a machine, not a function

An agent that installs a package, writes a file and runs a command needs a computer. A sandbox is one: isolated, started in about a second, thrown away when the work is done. One task cannot reach another, and an untrusted script cannot reach you.

A default sandbox is 2 vCPUs and 4 GiB at $0.018 an hour, billed per second. Build one with your toolchain, pause it, and fork it for each run: the fork starts with the dependencies already in place.

A sandbox also has a shell, a browser, SSH, open ports and a desktop, so the agent works the way a person would. See Sandboxes and the Sandboxes guide.

Let the agent reach your account

The MCP server gives an assistant your models, your usage and your sandboxes through one connection, so it can check a limit or start a machine without a bespoke integration.

What we do not do

We serve no embeddings endpoint, so a retrieval step needs an index of your own. See Knowledge search for how that fits together. We do not fine-tune, and we host no closed models.

What a security review will ask

An agent reads more of your data than a chat box does, so the questions get sharper: where the inference runs, where the sandbox runs, what either keeps, and who else touches the request.

Pin the region with the api.eu host and read munito-region off the response. We keep no prompts and no outputs by default, and we train on nothing you send. The DPA applies with no signature, and the sub-processor list names every company by region.

For the strictest work, run the platform inside your own cloud account or on-premises. See Enterprise.

Start