# Coding agents on open models and isolated compute

URL: https://munito.ai/solutions/coding-agents

Claude Code, Codex, OpenCode and five more, on open models in the EU with no proxy. Each agent gets a sandbox of its own, forked from a disk that already holds your toolchain.

## Your harness, open models, and a machine to run on

Coding agents are the workload that uses an API hardest: long context, many tool calls, hours of it, and a shell that has to exist somewhere.

Munito AI serves the models and the machines. You keep the harness you already like.

## Eight harnesses, each with a guide

Claude Code, Codex, OpenCode, Hermes Agent, pi, OpenClaw, DeepSeek Harness and Prime Agent each have a page in the documentation. See [the agent guides](https://docs.munito.ai/agents/claude-code).

**Claude Code needs no proxy.** The API answers the Anthropic Messages API on the same key as the OpenAI one, so you set a base URL and nothing else changes. Most providers make you run a translation layer; see [Compatibility](https://docs.munito.ai/compatibility).

```bash
export ANTHROPIC_BASE_URL="https://api.eu.munito.ai/inference"
export ANTHROPIC_AUTH_TOKEN="munitoai_..."
claude
```

## Give the agent its own computer

An agent that edits a repository, installs packages and runs a test suite needs a machine, and it should not be yours. A sandbox is isolated, starts in about a second, and goes away when the work is done.

| Size      | vCPUs | Memory | Per hour   |
| --------- | ----- | ------ | ---------- |
| `small`   | 1     | 1 GiB  | $0.0081    |
| `default` | 2     | 4 GiB  | $0.018     |
| `large`   | 4     | 8 GiB  | $0.036     |

Billed per second. A shell, a browser, SSH, open ports and a desktop come with it. See [Run an agent on a sandbox](https://docs.munito.ai/sandboxes/cloud-agents).

## Build the disk once, fork it for each run

The slow part of an agent run is rarely the model. It is `npm install`, the Docker pull, the dependency tree.

Build a sandbox with your toolchain, pause it, and fork it for every run. A fork starts with everything already installed, and no cache crosses the network. The same mechanism runs your CI. See [Persistence](https://docs.munito.ai/sandboxes/persistence) and [CI on sandboxes](https://docs.munito.ai/sandboxes/ci).

## The context is the cost

A coding agent re-sends the repository map, the instructions and the conversation on every turn. That prefix is most of the bill.

Cached input on DeepSeek-V4.1-Flash-1017 costs $0.006 per million tokens against $0.30 uncached, and the model reads a million tokens of context. GLM-5.3-Flash caches at $0.03 against $0.15. Price your agent on the cached rate, because that is what a long session actually pays.

## Your source code is the sensitive part

A coding agent reads the whole repository. That is the first thing your security team will say.

Send to `https://api.eu.munito.ai` and the request runs in the EU, with the region in the `munito-region` header. We keep no prompts and no outputs by default, and train on nothing you send. The DPA applies with no signature, at [legal.munito.ai/dpa](https://legal.munito.ai/dpa).

If the code may not leave your network at all, run the platform inside your own cloud account or on-premises. See [Enterprise](https://munito.ai/enterprise).

## What we do not do

We do not fine-tune a model on your codebase, and we host no closed model. If your team wants Claude or GPT weights, this is the wrong platform; if it wants the open models that now sit close behind them, at a posted price in a named region, it is the right one.

## Start

- [Connect Claude Code](https://docs.munito.ai/agents/claude-code)
- [Connect Codex](https://docs.munito.ai/agents/codex)
- [Run an agent on a sandbox](https://docs.munito.ai/sandboxes/cloud-agents)
- [Talk to an engineer](https://munito.ai/contact) about your team and your repositories