Understand a project
Ask ChatGPT to find files, read code and explain how the pieces fit together.
“Where is sign-in configured?”
Your local projects. Your ChatGPT.
Work from ChatGPT chat on your real local project, on Mac or Windows: explore and edit files, and run tests and commands in an isolated container with the permissions you approve. Connect local MCP servers too, without moving your projects to the cloud.
You choose the project and permissions. Read-only to start. Your files stay on your computer. The results ChatGPT requests are sent to OpenAI so it can answer. How your projects are protected
For chatgpt.com · Requires custom connector access Guided setup
Independent software, not affiliated with OpenAI
From chat to your files
Choose a folder in TunnelGPT, connect its tunnel to ChatGPT and start working in the conversation. Each project has its own connection.
On chatgpt.com, select your project connector and make a request. Your OpenAI account’s Secure MCP Tunnel carries it to TunnelGPT on your computer. TunnelGPT checks permissions and returns the permitted result. Your computer must be awake, online and running the tunnel. Requested results are sent to OpenAI so ChatGPT can respond.
Do more in the conversation
Read the original file when you need it and keep working with its context, without copying and pasting every snippet.
Ask ChatGPT to find files, read code and explain how the pieces fit together.
“Where is sign-in configured?”
Read notes and text, compare versions and prepare your next draft with information from the project.
“Summarize these notes and pull out the next steps.”
Enable editing when you need it and request the change. ChatGPT manages any required confirmation. A preview is optional: ask for one before applying a change.
“Prepare an update to the README.”
Project by project
Create an independent connection for each project. ChatGPT works with the folder you selected: it does not have to search across all your clients, repositories and subfolders. Switch connectors when you switch projects, with separate permissions and containers.
Your tasks keep their context
TunnelGPT automatically records references to file operations and project jobs alongside saved task goals and next steps. In the app, open Continue work for that connection to check jobs and copy or save a summary for another conversation.
As you work
Check jobs and copy or save the task’s continuation summary.
In another conversation
Resume this task and verify its latest results before continuing.
The app checks available operation receipts and job states when you resume. Your agent verifies the current files before making another change.
Task memory stays in private app storage. It does not save the whole chat or observe whether ChatGPT is still generating a response. Mac notifications are optional and can be changed in Settings.
How task memory worksNew in 0.1.10
Read and edit project files, then run commands or shell scripts with the execution permissions you approve. Use Rust and Cargo, Node.js, Python or media tools in the configured environment.
Prepare your project dependencies and run jobs in Docker. Jobs have no external network access. If you allow downloads, they are limited to dependency preparation. Chromium and HyperFrames rendering use this environment.
Advanced native execution runs in a macOS sandbox with approved toolchain folders. Additional interpreters and model files need readable paths you authorize. Outbound traffic is limited to loopback; listening services can be reachable from your LAN. Chromium rendering requires Linux.
Start with a private copy and review changes before publishing them to your project. Or explicitly allow commands to work directly in the project folder. Direct changes happen in place: keep a backup or version control because TunnelGPT cannot undo them.
Retrieve a job’s status, logs and results by its identifier. A chat response can end while a job on your computer is still running. Resume observation of that job instead of launching the same work again.
Video workflows depend on the project, installed tools, available disk space and any required models. Models are not all included in the app. Prepare dependencies and verify a short sample before a full render.
See setup and requirementsYour tools can stay local
Connect an MCP running on your computer and use its tools from ChatGPT through TunnelGPT. No VPS to rent, no server to move to the cloud. Your computer stays in control of the connection.
Add the address of a local MCP or choose an installed command. Browser tools, local apps and other integrations depend on the MCP you connect.
Discover the available tools and approve exactly which ones ChatGPT can use. Each MCP gets its own connection, which you can stop at any time.
Add the tunnel as a custom connector in ChatGPT. Ask it to use your local tools while your computer is awake, online and connected.
For macOS on Apple Silicon (M1 or later). Local MCPs keep their own permissions, including access to signed-in accounts or changes outside a project folder. Connect trusted servers and review their tools. ChatGPT connector access depends on your account.
Keep going wherever you are
Open ChatGPT on another computer or your phone and use your project’s connector when it is available in that interface. Your computer is still where the files live: it must stay awake, online and connected through TunnelGPT.
By voice, too. Ask ChatGPT to consult your project and keep talking in sessions that can invoke the MCP. Capability depends on the model and your account: check it during the trial.
Check compatibilityUntil December 2026, regular Chat has its own allowance, so you can keep working on your projects in Chat if your account still has Chat usage and connector access. From December 2026, OpenAI shares one allowance across ChatGPT and Codex. TunnelGPT connects your files; your OpenAI account determines the model and its limits.
Check it with the 3-day trial. Availability can change; TunnelGPT does not include or guarantee unlimited ChatGPT usage.Guided setup
You need your own OpenAI account with access to custom apps and Secure MCP Tunnels, a restricted credential, the exact tunnel ID, and Docker 24+ as the container runtime.
Security you can control
Start with read-only project access. Enable file editing and command execution separately, and choose the environment and folders those jobs can use. You control what is connected.
The file connection is scoped to your selected project. Excluded files and paths outside it are refused by file tools. Command execution and local MCPs have separate permissions, including any additional read folders you approve.
In the default mode, ChatGPT can read your files. Write tools are unavailable and the folder is mounted read-only.
Enable editing for the project you choose. File tools verify that the file has not changed, and the client handles required confirmation. Direct command execution has separate consent and can change project files without per-edit confirmation.
Commands and scripts require a separate execution grant. Choose an isolated Linux container or advanced native macOS execution, the workspace mode and any additional readable folders. Review these permissions before enabling execution.
The process that reads your files runs in a container without network access. The connector talking to OpenAI cannot mount or explore your project: it relays authorized responses.
Start or stop each connection in the app and check its status. Stopping it prevents the tunnel from serving new requests to that project.
The project file service and the connector use separate Docker containers. The file service uses --network none, a read-only system root, --cap-drop ALL, no-new-privileges and a private Unix socket (0600). File access starts read-only. Optional execution adds its own approved workspace and dependency mounts. These controls do not apply unchanged to native Mac jobs or external local MCP servers; review their separate permissions. The file service mounts its approved project with a bind mount.
Files stay on your computer. Content requested through the connector is sent to OpenAI to answer in ChatGPT; it does not pass through TunnelGPT’s commercial servers.
Privacy on your computer
How it works
Open technology. A focused desktop app.
The MCP core is available under the MIT License. Inspect the code, build your own integration and contribute improvements. TunnelGPT uses this same core in its commercial desktop application.
Project-scoped file access, confirmed writes, MCP messages and bounded local transport. Published under MIT with runnable examples and tests.
The desktop interface, guided setup, native integration, account and licenses are commercial. Future features may have a different scope; their availability in the public core is not guaranteed.
One product. One payment.
Try the connection, your projects and the way you want to work during your 3-day free trial, with no TunnelGPT usage quota. If it works for you and fits your needs, buy version 1. No subscription.
Evaluation
$0no card required
See how TunnelGPT fits the way you work, using a project of your own.
One trial per account. Ends 72 hours after your first successful connection.
Perpetual License
$9.97 USDone-time payment
Keep working with your projects in ChatGPT with a version 1 Perpetual License.
Not a subscription. Your TunnelGPT payment does not renew. Future major versions are not included.
Desktop app
Available for Mac Apple Silicon and Windows x64.
Supported: Apple Silicon Macs (M1 or later) and Windows 10 or 11, 64-bit (x64). Intel Macs: not supported yet.
Available for Mac Apple Silicon
Version 0.1.14 · Apple Silicon (M1 or later) · DMG. Signed with Developer ID and notarized by Apple.
TunnelGPT 0.1.14: Windows x64 joins macOS Apple Silicon, with file edits, commands and task memory over one ChatGPT connection.
Download for Mac · Apple Siliconcd415a804edf759f6a22630418a8b53635b502632918c8e33bf47918b4924193Available for Windows x64
Version 0.1.14 · Windows 10/11 x64 · installer. Not code-signed: Windows may show an unknown-publisher warning.
TunnelGPT 0.1.14: Windows x64 joins macOS Apple Silicon, with file edits, commands and task memory over one ChatGPT connection.
Download for Windows · x64The Windows installer is not code-signed, so Microsoft SmartScreen may show “Windows protected your PC”. Click “More info”, then “Run anyway”. You can check the file against the SHA-256 below.
4b3693fb4b421aaa0b1fe30bd9c39cbcbda78d98bb279c64382accdbba642af0Planned after the initial launch.
Download not available yetQuestions
Work on chatgpt.com with the files in a folder on your computer: find information, read code, review text and request changes. You start in read-only mode; editing requires enabling it and confirming changes. You can also connect a local MCP server and approve the tools you want to use from ChatGPT.
The whole folder is not synced. The MCP returns the results of requested tools, which do travel to OpenAI so ChatGPT can respond. Your original files stay on your computer. The connector cannot mount or explore your project.
The local app needs access to your selected project and uses your key to connect. Project content does not pass through our commercial servers, and your key is kept in your operating system’s secure storage.
You need a ChatGPT account with access to custom connectors, Secure MCP Tunnel permissions and Docker 24+. TunnelGPT does not include a ChatGPT subscription or third-party services. Docker Desktop is free for uses covered by its licence; other organizations need a Docker subscription.
TunnelGPT requires Docker 24 or later as its local container runtime. The project file service and the connector run in separate containers: the one that reads your project has no network access and starts read-only. Linux command jobs also run in a container without external network access; if you allow downloads, they are limited to dependency preparation. Docker Desktop licensing depends on your usage and organization.
Yes. Docker 24 or later is required for every connection, because the project file service runs in its own container. Advanced native execution only changes where approved commands run: those jobs use a macOS sandbox instead of the Linux container, with their own readable folders and network rules, so the container’s protections do not apply to them unchanged.
No. TunnelGPT is an independent product. It connects to OpenAI services using your account and permissions, but is not affiliated with, sponsored by or endorsed by OpenAI.
Yes, when you explicitly enable command or shell-script execution for the project. Linux jobs run offline in Docker; advanced native jobs use a macOS sandbox and approved toolchain folders. A private copy and direct project work have different change controls. Direct commands can modify files without per-edit confirmation. Local MCP servers can have broader permissions; approve their tools separately.
You choose which project folders to connect. Each project starts read-only; use Allow writing to let ChatGPT propose changes that you preview and confirm, and to allow remote development jobs. You approve each local MCP tool. For commands, choose an isolated Linux container (recommended) or advanced native execution on this Mac in a macOS sandbox, and work on a private copy or directly in the project folder. You can stop any connection at any time.
Yes. Each folder has its own profile and an independent tunnel. In ChatGPT, choose the connector for the project you want to work on. Connecting one folder does not grant access to the others.
Yes. It must stay awake, online and running Docker and the TunnelGPT connection. You can access it from another compatible ChatGPT session; the tunnel does not wake your computer or keep it connected when it shuts down.
No. The price is $9.97 USD, paid once for a version 1 Perpetual License. Use it on up to 2 devices and release one from your account to switch. Future major versions are not included. Check current availability before buying.
The trial gives you 3 free days of unlimited TunnelGPT use: 72 consecutive hours from your first successful connection, including disconnected time. ChatGPT limits depend on your account. After purchasing, you have 14 calendar days to request a refund at support@tunnelgpt.com.
You can work by voice in sessions that support your MCP; we have tested continuous voice. Mobile and voice availability, and the complexity of what the model can do, vary by account and interface. OpenAI’s public documentation may differ from features being rolled out. Use the 3-day trial to check your use case before buying.
Until December 2026, yes, if your account still has Chat usage and connector access available: Chat has its own allowance, so you can continue working with local files in the conversation. From December 2026, OpenAI shares one allowance across ChatGPT and Codex, so this no longer saves usage. This does not transfer tokens or extend your plan. OpenAI determines limits, models and permissions, which may change. Verify your account and workflow during the free trial.
The tunnel can connect to an MCP on a remote machine. To use TunnelGPT, that machine must run a compatible version of the app and Docker, have access to the folder and stay awake and online. The current distribution is a desktop app; it does not include hosting or a Linux server edition.
Connect your first project and decide what ChatGPT can do with it. Your conversation, your computer and your permissions.