Article 14 in the AgentKit v3 onboarding series. About a 5 minute read. Article 10 briefly introduces CloudHarness; this one goes deep on how to use it.
My starting idea was pretty lazy: ChatGPT is great at chatting about code, but it has no hands and feet. Codex has hands and feet, but the quota is limited. So why not give ChatGPT its own set of hands and feet through MCP? The model thinks inside ChatGPT, while reading files, editing code, running commands, and Git happen in an external harness. That is Cloud Harness MCP.
A disclaimer up front to avoid misunderstanding: I have no measurements proving this "doubles your quota" or gives "unlimited GPT". Limits change by plan and by time. All I can say for sure is that in my experience, ChatGPT still had its own usage left when Codex had hit its cap. Do not treat that as a promise.
1/ What is it?
An authenticated MCP server over HTTP, running just for you (or a team that trusts each other). The flow:
The client (ChatGPT or another MCP client) calls into an authenticated API layer.
The API hands the job to a trusted runner.
The runner executes commands in a Docker executor, non-root, with time limits and a fenced-off workspace.
GitHub credentials live in a separate broker, not permanently inside the container that runs code.
Each working session is a workspace: the agent calls workspace_open with a repo, records the workspaceId, works through the tools (files, symbol search, shell, Git, tasks, subagents, skills, hooks, depending on config and permissions), then calls workspace_close.
2/ When to use it, when not to
| Situation | A good fit? |
|---|---|
| Brainstorming architecture, reviewing, planning with a strong model in ChatGPT, then letting it touch the repo itself | Yes |
| You want the agent to work on another machine, so you do not run commands on your laptop | Yes |
| One person or a small team that trusts each other | Yes |
| Multi-customer infrastructure, strangers sharing one instance | No |
| You need a "sandbox against adversaries" | No |
The reason for the last line: the executor shares a kernel with the host. This is a single trust domain, not a sandbox for multiple parties who do not trust each other.
3/ Using it, step by step
Step 1: deploy and turn on authentication. Follow the project docs at docs.harness.agentkit.best. Read the security model and check the secrets before running, don't let the agent install it for you and only read afterwards.
Step 2: connect a client. Article 10 lists two live endpoints:
Clients that support OAuth:
https://harness.zuey.me/mcpClients that can only send a static header:
https://api.harness.zuey.me/mcp, using an API key from the dashboard
Not every ChatGPT interface accepts remote MCP configuration. Check your plan and settings first.
Step 3: tell the agent to open the repo. It calls workspace_open, then reads the repo's AGENTS.md and architecture docs.
Step 4: work. Find symbols, edit files, run tests, create commits. Pushing and opening a PR should only be allowed after you approve the action.
Step 5: clean up. Close the workspace, then check for any leftover processes or tasks still running.
Run locally without a VPS (stdio mode, POSIX only; on Windows use WSL):
cloud-harness-mcp --transport stdio --workspace /duong/dan/repoNote: stdio is for clients that can run MCP locally. ChatGPT usually connects through remote MCP, so don't expect pasting the command above into ChatGPT to work.
4/ Status by part
| Part | Label | Notes |
|---|---|---|
Remote MCP over HTTP, workspace_open/workspace_close |
✅ beta | Running live |
| Local stdio mode | ✅ | POSIX; not an OS-level sandbox |
| Shared use across customers who do not trust each other | 🗺️ | Not a current goal |
| ChatGPT allowing this kind of usage | 🧪 | Depends on plan and OpenAI terms. I cannot claim OpenAI approves it |
The latest stable release per article 10 is 0.76.2, MIT-licensed source: bestagentkits/cloud-harness-mcp.
Wrap-up
1/ Cloud Harness is MCP hands and feet for a model: the model thinks in ChatGPT, the real work runs in the harness.
2/ Use it for one person or a team that trusts each other. It is not a multi-customer sandbox.
3/ Quota and terms depend on the plan you pay for. MCP is a legitimate protocol, but you are responsible for complying with the terms of service.
4/ Read the security model, and protect your API key and GitHub permissions before running.
Sources
Cloud Harness MCP: turn ChatGPT into a coding environment (10/10/2026, status as of the source code that day)