Coding Harness — Run Any AI Coding CLI in a Hosted Cloud Box

What Is a Coding Harness?

A coding harness is the program that drives an AI model through real work in a terminal — reading your files, editing them, running builds, and reacting to what the tests say. The model on its own only produces text. The harness is what turns that text into commands, feeds the results back, and keeps the loop going until the job is done. Claude Code, Codex, OpenCode, Aider, OpenClaude, Zero, Pi, and DeepSeek Harness (dsh) are all coding harnesses. They differ in which models they drive, how much they do before asking you, and how they handle large codebases — but they are the same category of tool.

Two words worth keeping apart, because they get used interchangeably and mean different things. The harness is the agent. The workspace is the box it runs in — the container, its disk, its terminal, the thing that keeps your code between sessions. You need both: a harness with nowhere to run is just a package on npm, and a box with no harness in it is an empty Linux machine.

That is what this page is about: running any of the eight harnesses in a hosted box, so you do not have to install Node, Python, a coding CLI, and its dependencies on your laptop first. You open the box in your browser, pick a harness from a dropdown, and work in a real terminal against a persisted /workspace directory.

It is a different product from a chatbot. OpenClaw Launch also hosts AI agents that talk to you on Telegram, Discord, Slack, and the web — but a coding box has no messaging channel and no chat gateway. It is a developer environment: an agentic CLI sitting in a terminal, editing files, running builds, and executing tests on your code. Think of it as a cloud-hosted home for tools like Claude Code and Aider that you would otherwise run on your own machine.

The CLIs You Can Run

Every Coding Workspace ships with these terminal coding agents pre-installed. You switch between them from the workspace card — they all operate on the same /workspace code, so you can use whichever one fits the task.

Because they all live in the same box, a Coding Workspace is a low-friction way to try the CLIs side by side and decide which one you prefer — without a separate local install for each.

How Login Works — and Why We Don't Inject Your Keys

Each CLI owns its own authentication. Claude Code and Codex use their native sign-in — a Claude or ChatGPT subscription OAuth flow, or a pasted API key. OpenCode uses opencode auth login. Aider reads a provider key from the environment or a flag. Pi reads a provider key from the environment too, or signs in with /login. Zero picks up an OpenAI or Anthropic key from the environment the same way, with one exception worth knowing: it will not use an OpenRouter key it merely finds there — OpenRouter has to be registered with zero providers add openrouter first, or Zero reports no provider configured while a perfectly good key sits in the environment. You authenticate each tool the same way you would locally, straight from the terminal.

One deliberate design choice is worth calling out: the workspace never places your saved provider API keys in the environment under the names a coding CLI reads. An ambient API key would silently override a subscription login and bill your pay-per-token account instead of your flat-rate Claude or ChatGPT plan, so your native sign-in stays in control of billing. The one key that does travel into the box is carried under a private name no CLI looks for, and only Zero — which cannot be given a key any other way — is pointed at it.

To make a one-time sign-in survive the disposable-box churn, the workspace persists the agent home directory — where every CLI stores its login and config — alongside your code. You sign in once, and the login carries across resets.

Persisted Workspace, Reset, and Wipe

A Coding Workspace is disposable by design but not forgetful. Two things persist independently of the container: your code under /workspace, and your CLI logins and config in the agent home directory. The container around them is throwaway, which is what makes the box safe to recreate at any time.

• Reset — recreates the container against the same storage. You get a clean box, but your code and your CLI sign-ins are intact. Use this when the environment gets into a weird state and you want a fresh start without losing work.
• Wipe — drops the storage first, then recreates. This is a true clean slate: code and logins are both gone. Use it when you want to hand the box to a new project with nothing carried over.

Why Run a Coding CLI in the Cloud?

No Local Install or Toolchain Setup

Installing an agentic coding CLI and getting its runtime and dependencies right is its own chore. A hosted workspace ships with every agent and what they need already installed in a pinned image, so you skip the "works on my machine" setup and start coding in seconds from any device with a browser.

Runs on a Server, Not Your Laptop

The workspace runs in a server-side container, so the box itself stays up regardless of whether your machine is awake or online. You can open the terminal from any device and pick up where you left off — your code and environment are already there, not tied to one computer.

Isolated From Your Own System

An agentic CLI executes commands. Running it in an isolated cloud container keeps that execution away from your personal files, SSH keys, and system. Each workspace runs in its own hardened container — no Docker socket access, no new privileges, dropped Linux capabilities, and resource caps — so an agent stays inside its box.

Try Multiple Agents Without Installing Any of Them

Comparing Claude Code, Codex, OpenCode, Aider, OpenClaude, Zero, Pi, and DeepSeek Harness locally means a separate install and a separate auth flow each to maintain. In a Coding Workspace they are already there; you just switch and go.

How to Start a Coding Workspace

1. Open the deploy page — Go to openclawlaunch.com, sign in, and choose the Coding Workspace option on the deploy card. A free trial is available, so you can try it before subscribing.
2. Deploy the box — OpenClaw Launch provisions a hardened container from the pinned coding image, attaches your persisted /workspace and home storage, and starts it. The workspace shows up as a card on your dashboard.
3. Open the terminal — Click Open Terminal to get a real browser terminal running as the box's non-root user. Your code lives in /workspace.
4. Pick an agent and sign in — Choose Claude Code, Codex, OpenCode, Aider, OpenClaude, Zero, Pi, or DeepSeek Harness, then complete its sign-in or provider setup once. It persists across resets.
5. Start coding — Point the agent at your project and let it read, edit, build, and test. Use Reset for a fresh box that keeps your work, or Wipe for a clean slate.

A Hosted Coding Harness vs Running the CLI Locally

Local installs are great when you are already at your own machine and happy to manage toolchains. A hosted workspace wins when you want zero setup, isolation from your own system, access from any device, or a quick way to compare agents.

Related Guides

• Claude Code with OpenClaw — using Claude Code as a model backend for a chat bot
• Codex with OpenClaw — Codex CLI setup and OAuth
• Aider with OpenClaw — running Aider in the OpenClaw stack
• OpenCode with OpenClaw — OpenCode integration
• Pi coding agent — setup, models, non-interactive runs, and sandboxing
• OpenClaw Launch hosting plans — pricing and feature details

Frequently Asked Questions

What is a coding harness?

A coding harness is the program that drives an AI model through real work in a terminal — reading your files, editing them, running builds, and feeding the results back to the model until the job is done. Claude Code, Codex, OpenCode, Aider, OpenClaude, Zero, Pi, and DeepSeek Harness (dsh) are all coding harnesses. The harness is the agent; the workspace is the box it runs in. On OpenClaw Launch the box comes with Claude Code, Codex, OpenCode, Aider, OpenClaude, Zero, Pi, and DeepSeek Harness ready to use, with a browser terminal and a persisted workspace directory, so you can run any of them without installing anything locally.

Can I use my Claude or ChatGPT subscription instead of an API key?

Yes. Claude Code and Codex support their native subscription sign-in inside the workspace, and the box deliberately never places a provider API key in the environment under a name those CLIs read, so an ambient key cannot silently override your subscription login and bill a pay-per-token account. Your sign-in stays in control of billing, and it persists across resets.

Does my code survive if I reset the workspace?

Yes. A reset recreates the container but keeps both your code in /workspace and your CLI logins. A wipe is the one that clears everything — it drops the storage for a true clean slate. Use reset for a fresh environment that keeps your work, and wipe to start over completely.

Is it safe to let an AI coding agent run commands in the workspace?

Each workspace runs in its own hardened container — no Docker socket, no new privileges, dropped Linux capabilities, and resource limits — and the agent runs as a non-root user. That isolation is a key reason to run an agentic CLI in a hosted box rather than directly on your own machine, where it would have access to your personal files and credentials.

How much does a Coding Workspace cost?

Lite is $3 for your first month and $6/month after that; Pro is $20/month. A free trial is available so you can try a workspace before subscribing. You bring your own model access through each CLI's native sign-in or API key.