Yes, Zero Risk fits your safety constraint better because it does not read or operate the ChatGPT page and leaves the send action for your confirmation; the trade-off is that continuous automatic sending is lost.

• The README explicitly says, “Zero Risk does not read or operate the ChatGPT page.”
• It requires you to choose the model, select the Codex Zero Risk connector, paste the prepared prompt, and confirm Sent in the launcher.
• The mode table lists Sending messages as “Paste and send manually,” while Local Codex tools remain available through a separate MCP connector.
• Compared with Full harness, it reduces browser automation and accidental-send risk, but it cannot provide the same continuous autonomous execution experience.

It therefore fits workflows requiring human confirmation for every Web submission while retaining an MCP tool channel. If you also prohibit all tool calls, you should not enable that connector.

Yes, because your goal is to reuse ChatGPT Web models rather than connect local tools; Browser-only provides automatic sending and a Codex model entry without granting local access.

• The README says the project can use ChatGPT Web models available on the account, including Pro, with separate Web usage limits.
• Browser-only sends messages automatically, while Local Codex tools are marked No, matching your constraint.
• Models exposed by Plus or Pro are detected by the launcher. Pro versions have separate entries, and reasoning strength is selected through Effort.
• After installation, you must sign in, run the browser smoke test, install models, and restart Codex once; otherwise the model entry may not appear as expected.

If you later need to edit the current repository or run commands, Browser-only is no longer sufficient as a full tool-enabled agent.

Yes. The README explicitly provides a Luna / Think path for Free and Go accounts, so a reasoning selector or Pro entitlement is not required.

• The Available models section states: Free/Go → Luna / Think.
• The README says accounts with reasoning controls instead receive Instant–High, with Extra High and Pro when available.
• The launcher detects which models the account can actually use, so the available entry depends on account state rather than a manually forced Pro configuration.
• After installing models, you must restart Codex; in automatic mode, choose a model whose name ends in (Web).

This means the project can improve Codex access for Free or Go users, but it cannot upgrade the account to Pro or guarantee the quota and long-term availability of Luna / Think.

Yes, provided that you can complete the MCP and ChatGPT Developer Mode setup; Full harness is the mode that explicitly connects the current Codex task to local tools.

• The Full harness documentation says MCP connects ChatGPT to the current task’s files, terminal, tools, and approvals.
• The mode table marks Full harness as Automatic for sending and Yes for Local Codex tools, while Browser-only is explicitly No.
• The project insight requires an outbound tunnel through the official OpenAI tunnel-client, plus Codex Native2 configuration, workspace permissions, and approval behavior.
• This is more than text forwarding: task-linked conversations allow context to remain associated with the Codex task.

It therefore fits interactive coding, but it is not an unsupervised local executor. The chain can still fail if the workspace or administrator policy blocks MCP, writes, or tool calls.

It depends. If the workspace blocks Developer Mode, MCP, or writes, Full harness is not suitable, although Browser-only may still work for tasks that do not need local tools.

• The README bases Full harness tool access on MCP and tells users to open MCP setup for coding with tools.
• The project insight adds that Full harness requires ChatGPT Developer Mode, workspace permissions, and potentially administrator approval.
• Browser-only explicitly has Local Codex tools set to No, so it cannot bypass workspace policy to access files, terminals, or approvals.
• Zero Risk avoids automatic ChatGPT-page operation, but it still relies on a separate MCP connector for local tools; it is not a workaround for a workspace prohibition.

The practical decision is therefore: use Browser-only in a managed workspace if local tools are unnecessary; before using any tool-enabled mode, confirm that Developer Mode, MCP applications, and the required workspace operations are approved.

It depends. Those three platforms are within the project’s stated coverage, and the application includes its browser and runtime, but platform signing, first-run warnings, and unsupported architectures may still block deployment.

• The README lists macOS Intel, while the project insight identifies macOS 13+ arm64/x64, Windows x64, and Linux x64 as the main supported package targets.
• The Get started section says the app includes its browser and runtime, so separate Chrome, Node, or Bun installation is not required.
• Installers select the platform and architecture, verify published checksums, and preserve the ChatGPT profile and launcher settings.
• The project data identifies TypeScript as the main language and v6.1.2 as the latest release; the project insight also states that release packages are not platform-signed, which may trigger Gatekeeper or SmartScreen.

The target platforms and no-extra-runtime requirement therefore match well. If your organization requires signed installers or a different architecture, the README provides no sufficient guarantee.