---
title: "Harnesses and models"
description: "The three agents Skyworker runs, how to choose one, and where models and keys come from."
---

> Documentation Index
> Fetch the complete documentation index at: https://skyworker.system3inference.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Harnesses and models

A harness is an agent program that runs a session. Skyworker has three. All three are in the workspace image.

| Harness | Name in `--harness` | Notes |
| --- | --- | --- |
| OpenCode | `opencode` | The default. Your local OpenCode can open its sessions. |
| Claude Code | `claude-code` | Runs with no permission prompts. |
| Codex | `codex` | Runs with no permission prompts. |

Choose one for each task:

```sh
skyworker run "Fix the flaky test in auth" --harness claude-code
skyworker run "Add input checks to the signup form" --harness codex
```

## What differs between harnesses

- The workspace is the sandbox for Claude Code and Codex. They ask for no permission.
- The OpenCode client of your machine cannot open a session of Claude Code or Codex. Use the `tasks` commands or the phone client. See [Use Skyworker from a phone](https://skyworker.system3inference.dev/docs/guides/phone/).
- The session id of Claude Code and Codex is a UUID. Name its task as `<workspace>:<id>`, or pass `--workspace`.
- A question or a permission request from Claude Code or Codex to a person is not built. Those two harnesses never reach the state `waiting`. See [Limits](https://skyworker.system3inference.dev/docs/limits/).

## Models and keys

No model key is needed. The Worker has an AI binding. A harness in a workspace calls a host that the platform intercepts. The Worker runs the request on your account.

OpenCode has two providers:

| Provider | Models | Billed to |
| --- | --- | --- |
| `cloudflare-workers-ai` | Workers AI models (`@cf/...`) | Workers Paid usage |
| `cloudflare-ai-gateway` | Third-party models | AI Gateway credits |

The default OpenCode model is `cloudflare-workers-ai/@cf/zai-org/glm-5.3`. Change it with the `model` setting, or per task with `--model`. See [Configuration](https://skyworker.system3inference.dev/docs/reference/configuration/).

Claude Code and Codex use the binding when the Worker holds no key for them. On the binding the defaults are `anthropic/claude-sonnet-4.5` for Claude Code and `openai/gpt-5.1` for Codex. `--model` takes a model name of that harness, in the names of the binding, for example `anthropic/claude-haiku-4.5`.

To use your own provider account, set its key as a Worker secret. A key wins over the binding:

```sh
npx wrangler secret put ANTHROPIC_API_KEY --name skyworker
npx wrangler secret put OPENAI_API_KEY --name skyworker
```

The key never enters the workspace. The workspace sends a placeholder, and the Worker adds the real key on the way out. See [Security](https://skyworker.system3inference.dev/docs/concepts/security/).

A provider that you log in to from inside a workspace is different. Its credential is written to the workspace disk and into every snapshot after. Use the Cloudflare providers or a Worker secret instead.

Third-party models bill to your account through AI Gateway. The AI Gateway is `default` unless the `aiGateway` setting names another.

Source: https://skyworker.system3inference.dev/docs/guides/harnesses/index.mdx
