Comparison

ChatGPT desktop app vs TunnelGPT for local projects

If you use the ChatGPT desktop app, Work and Codex can already open a local folder, edit files and run commands as part of your plan, and for many people that is the simplest choice. TunnelGPT is for a different case: working on a Mac project from regular ChatGPT Chat on chatgpt.com, including from another computer or your phone when the connector is available there, with one connection per project, a read-only start and commands in an offline container. It also helps when your Work or Codex allowance runs out: our team tested continuing in Chat on Plus and Pro.

Checked on 25 October 2026. OpenAI's pages carry no date stamps and change often, so each claim below links to the official source. TunnelGPT is independent software, not affiliated with OpenAI.

What the ChatGPT desktop app already does

The ChatGPT desktop app for macOS, Windows and Linux has three modes: Chat, Work and Codex. According to OpenAI, Work and Codex can open a local folder, edit files and run commands, following the sandbox and approval settings you choose (Codex app, get started with Work). Both are included in Free, Go, Plus, Pro, Business, Edu and Enterprise (pricing).

The desktop app and the Codex CLI can also use local stdio MCP servers (Codex MCP documentation). Local MCP tools in desktop Chat sessions have an open bug report, so check its status before relying on them in Chat.

One detail matters for planning: “ChatGPT Work usage inside ChatGPT uses the same pricing, credits, and usage limits as Codex” (OpenAI pricing). Work and Codex draw on the same allowance.

Can ChatGPT on the web access my files?

Not on its own. OpenAI says Work on web and mobile cannot directly access files on your computer (OpenAI Help). ChatGPT on the web reaches remote tools through plugins and custom MCP servers, and it rejects localhost URLs as “Unsafe URL”. The documented route to a private or localhost MCP server is OpenAI's Secure MCP Tunnel, which needs a tunnel ID from the OpenAI Platform and an API key with Tunnels Read + Use.

TunnelGPT packages that route for Mac projects. It runs OpenAI's official tunnel client, pinned and hash-checked, gives each project its own connection, and walks you through creating the tunnel and a restricted key. It does not create them for you. No TunnelGPT relay sits in the data path. Step by step: connect ChatGPT to local files.

Which ChatGPT surface can do what

When the desktop app is the better choice

• You work at the computer where the project lives, and Work or Codex already fits how you work.
• You want everything included in your plan, with no extra app, Docker, tunnel or API key to set up.
• You use Windows or Linux: the ChatGPT desktop app runs there, while TunnelGPT currently needs a Mac with Apple Silicon (M1 or later).
• You rely on local stdio MCP servers in Codex, which the desktop app supports directly.
• You want to steer Codex from your phone: Remote connections pair a phone with a desktop host, using Codex access and Codex usage.
• Your organization does not let you create a Secure MCP Tunnel or a restricted key in the OpenAI Platform, which TunnelGPT requires.

When TunnelGPT fits

• You want to work from chatgpt.com. You prefer regular ChatGPT Chat in the browser, or you are on another computer, and want a Mac project available there as a connector.
• Your phone or a voice session should reach the project. When the connector is available in that interface, you can use it from ChatGPT on a phone, including voice sessions that can invoke the MCP. Capability depends on model and account.
• Your Work or Codex allowance ran out. Our team tested continuing in regular ChatGPT Chat on Plus and Pro after using up Codex and Work, with reading, editing and commands. If your account still has Chat usage and connector access, you can keep working there until December 2026; from December 2026 OpenAI shares one allowance across ChatGPT and Codex, so this no longer saves usage. This is not a promise for every account, and availability can change. See Codex limit reached.
• You want strict per-project boundaries. Each project gets its own profile and tunnel, starts read-only, and needs separate grants to edit or run commands. Writes check the file's expected hash, re-read to confirm and are journaled with receipts.
• You want commands kept apart from your Mac. Commands run in an isolated Linux container in Docker with no external network, or in an advanced macOS sandbox. By default a job works on a private copy that you review before publishing.
• You want long jobs you can check later. Jobs keep running on the Mac if the chat stops, and a new conversation can read their status, logs and results.
• You want local MCPs in Chat on the web. Add them by HTTP address or command and approve which tools ChatGPT can use, with no VPS or cloud hosting.

What you need: a Mac with Apple Silicon, Docker 24+ running, a ChatGPT account that can add custom MCP servers, and permission in your OpenAI Platform organization to create a Secure MCP Tunnel and a restricted key with Tunnels Read + Use. The price is $9.97 USD once for a Version 1 Perpetual License, after a 3-day trial with no card required. See pricing.

What TunnelGPT does not change

• It does not add OpenAI usage or model access; your plan's limits still apply.
• It is not local-only AI. Original files stay on the Mac and do not pass through TunnelGPT's servers, but the tool results ChatGPT requests are sent to OpenAI so it can answer.
• It does not resume an interrupted ChatGPT response or restore a chat automatically. Task memory saves a goal and next step that you can pick up in a new conversation.
• It needs the Mac awake, online and connected while you work.

Using both together

The two are not exclusive. One way to combine them: use Work or Codex in the desktop app at your desk, where it is included in your plan, and keep the same project connected through TunnelGPT for regular Chat on chatgpt.com when you are on another computer, or on your phone. Until December 2026 this also helps when you are out of Codex and Work usage; from December 2026 Chat and Codex share one allowance.

Avoid having both edit the same files at the same time. TunnelGPT's private-copy mode lets you review a job's changes before publishing them, and its writes check the file's expected hash before changing it. Before switching, ask the agent to save the task's goal and next step so the next conversation starts from a clear point.

Can the ChatGPT desktop app access local files?

Yes. In the desktop app, Work and Codex can open a local folder, edit files and run commands under your sandbox and approval settings. That is included in Free, Go, Plus, Pro, Business, Edu and Enterprise.

Is a ChatGPT Work local folder the same as a connector?

No. Work opens the folder directly in the desktop app and uses Codex usage. A connector is a custom MCP server that ChatGPT calls from Chat; for a private server on your Mac, it goes through OpenAI's Secure MCP Tunnel. TunnelGPT provides that connector for one project at a time.

Does TunnelGPT use my Codex or Work usage?

In our testing on Plus and Pro, work through TunnelGPT in regular Chat kept going after Codex and Work usage was used up, so it drew on Chat usage. OpenAI states separately that Work and Codex share limits. From December 2026, OpenAI moves Plus, Pro and Business plans to one allowance shared by ChatGPT and Codex, measured in tokens, with only Instant chat not drawing from it (OpenAI help center); from then on, work through TunnelGPT in Chat uses that same allowance. Your limits depend on your plan, and OpenAI can change them.

Do I need the desktop app to use TunnelGPT?

No. TunnelGPT runs on your Mac and connects your project to ChatGPT on chatgpt.com through OpenAI's Secure MCP Tunnel. To compare TunnelGPT with open-source tools, see TunnelGPT vs open-source ChatGPT bridges. To start, follow the setup guide.