Skip to main content
Glama

On Glama

Glama hosted release v0.1.0 is published. This restricted, read-only documentation sample exposes five genuine MCP tools against two bundled public Markdown files. It does not access visitors' computers, private folders, Chrome sessions, shell, OpenAI tunnels or credentials.

View the Glama listing · Hosted release details · Installation and security differences

The full MCP Bridge remains a separate local desktop application. Install it on Linux or Windows, approve specific folder roots and optional terminal or browser access, and connect using your own authorized OpenAI Secure MCP Tunnel. Glama's hosted sample is for protocol demonstration and tool inspection, not remote control of a desktop.

Related MCP server: agent-mcp-gateway

What is MCP Bridge?

MCP Bridge is a desktop app for Linux and Windows that keeps powerful local tools on your own computer while making them available to ChatGPT through OpenAI Secure MCP Tunnel.

It can manage three independent connections (the third requires Meta Manager):

Connection

What it does

Local MCP

Browse/search approved folders, read files, optionally create/edit files, and optionally run terminal commands.

Web MCP

Control your existing Chrome session through Chrome DevTools MCP: tabs, navigation, clicks, typing, screenshots, console/network inspection, and optional vision tools.

Meta Manager MCP

Optional connection to a separately installed and running Meta Manager local MCP service.

Your files, browser session, and terminal remain on your machine. MCP Bridge only exposes the roots and capabilities you explicitly allow.

Highlights in v0.2.1

  • Fresh product-level desktop UI

  • Dedicated Bridge / Gateway app identity and production icon

  • Proper Linux taskbar/window identity — no longer grouped under ChatGPT

  • System tray controls for Local MCP and Web MCP

  • Closing the window minimizes instead of terminating active tunnels

  • Automatic Linux mounted-drive discovery

  • Friendly labels for external SSDs and mounted volumes

  • Read/write validation before a mounted root is exposed

  • Existing encrypted Runtime API key survives upgrades

  • Built-in cross-platform update system

  • SHA-256 verification before an update can be installed

  • Linux .deb, Linux AppImage, and Windows NSIS .exe releases

  • Local MCP and Web MCP remain completely independent

Architecture

                         OpenAI Secure MCP Tunnel
                                  |
                  +---------------+---------------+
                  |                               |
              Local MCP                        Web MCP
                  |                               |
          approved local roots              Chrome DevTools MCP
         /        |        \                       |
      files    folders    shell               existing Chrome
                  \                               /
                   +--------- MCP Bridge --------+
                             Linux / Windows

No public local MCP port needs to be exposed directly to the internet.

Quick start

1. Install MCP Bridge

Linux

Current packaged releases target Debian / Ubuntu / Linux Mint on amd64/x86_64.

curl -fsSL -H "Accept: application/vnd.github.raw+json" \
  "https://api.github.com/repos/haseebgb92/ChatGPT-MCp/contents/install-linux.sh?ref=main" | bash

If Chrome/Chromium is already installed:

curl -fsSL -H "Accept: application/vnd.github.raw+json" \
  "https://api.github.com/repos/haseebgb92/ChatGPT-MCp/contents/install-linux.sh?ref=main" | SKIP_CHROME=1 bash

Or download the latest .deb from GitHub Releases:

sudo apt install ./MCP-Bridge-*-linux-amd64.deb

Windows

Open PowerShell:

Set-ExecutionPolicy -Scope Process Bypass -Force
irm "https://raw.githubusercontent.com/haseebgb92/ChatGPT-MCp/main/install-windows.ps1" | iex

Or download the latest Windows .exe from GitHub Releases.

2. Create two OpenAI tunnels

Open:

https://platform.openai.com/settings/organization/tunnels

Create separate tunnel IDs for Local MCP and Web MCP.

Local Files
Chrome

Use different tunnel IDs when both services will run at the same time.

3. Create a Runtime API key

Open:

https://platform.openai.com/settings/organization/api-keys

For least privilege, use a restricted key whose principal has:

Tunnels: Read
Tunnels: Use

Do not use an OpenAI Admin API key as the long-running Runtime API key.

When Remember API key securely is enabled, MCP Bridge stores it through Electron safeStorage / operating-system credential encryption instead of normal settings.

4. Configure MCP Bridge

  1. Enter the Runtime API key.

  2. Enter the Local tunnel ID.

  3. Enter the Web tunnel ID.

  4. Add one or more approved folders.

  5. Choose read-only or read/write access per root.

  6. Enable file writing only if needed.

  7. Enable terminal access only if needed.

  8. Enable Browser vision tools if needed.

  9. Optionally start MCP Bridge at OS login.

  10. Start Local MCP, Web MCP, or both.

5. Add the connections in ChatGPT

In ChatGPT, enable Developer mode and add MCP connections using Tunnel as the connection method.

Local connection:

Name: Local Files
Connection: Tunnel
Tunnel: <your Local tunnel ID>

Browser connection:

Name: Chrome
Connection: Tunnel
Tunnel: <your Web tunnel ID>

Start the matching service in MCP Bridge before refreshing/scanning its tools in ChatGPT.

Mounted drives on Linux

MCP Bridge automatically discovers mounted filesystems under common Linux mount locations such as:

/mnt
/media/<user>
/run/media/<user>

Instead of making you work with an opaque mount path such as:

/mnt/289E678E9E6752FC

the UI can show the volume label, filesystem, size, and actual mount path.

Before a mounted root is exposed, MCP Bridge resolves the real path and confirms that the current user can read it. Read/write roots are also checked for write access.

This makes external SSDs and secondary drives first-class without weakening the Local MCP path-security model.

Desktop experience

  • Dedicated app and tray icon

  • Separate Linux WM_CLASS / desktop identity

  • Click X to minimize instead of killing active tunnels

  • Tray menu to reopen the app

  • Start/stop Local MCP from the tray

  • Start/stop Web MCP from the tray

  • Restart running services

  • Explicit Quit when you want the bridge terminated

  • Optional launch at login

  • Optional automatic Local/Web startup

  • Runtime activity logs

Built-in updates

Starting with v0.2.1, MCP Bridge can update itself from GitHub Releases.

Update flow

  1. Check the latest GitHub Release.

  2. Compare semantic versions.

  3. Select the correct installer for the current OS and architecture.

  4. Download the installer locally.

  5. Download the matching .sha256 file.

  6. Verify the installer before installation is enabled.

  7. You explicitly choose Install & restart.

Automatic checks happen shortly after launch and then periodically. They can be disabled in Settings.

Linux

download → SHA-256 verify → pkexec → apt install → relaunch MCP Bridge

The system elevation dialog is used. MCP Bridge does not store your sudo password.

Windows

download NSIS .exe → SHA-256 verify → launch installer → MCP Bridge exits

Updates are not silently installed.

Local MCP

Files and folders

  • Multiple approved roots

  • Read-only/read/write mode per root

  • Mounted Linux drive support

  • List files and directories

  • File metadata

  • Read UTF-8 text files

  • Recursive filename/content search

  • Create folders

  • Create or replace text files

  • Real-path and traversal protection

  • Symlink escape protection

Shell

Terminal execution is optional.

Restricted shell supports normal development work while blocking several obvious destructive/system-level commands.

Examples:

npm install
npm run build
git status
git diff
git pull
git push
python script.py
adb devices

Restricted shell is a guardrail, not a complete OS sandbox.

Full shell executes Bash on Linux or PowerShell on Windows with the current user's privileges.

Treat Full Shell as local-user-equivalent code execution. Enable it only for a machine and MCP connection you trust.

Web MCP / Chrome

Web MCP uses chrome-devtools-mcp with your existing Chrome session, so ChatGPT can work with tabs and authenticated sessions already open on your computer.

Capabilities include:

  • list/select/open/close tabs

  • navigation

  • click, hover, typing, forms, and keyboard input

  • screenshots and accessibility snapshots

  • console and network inspection

  • performance tooling

  • optional coordinate/vision tools

  • optional usage-statistics opt-out

Chrome setup

Open:

chrome://inspect/#remote-debugging

Enable remote debugging for the Chrome profile you want MCP Bridge to use.

When Browser vision tools are enabled, MCP Bridge launches Chrome DevTools MCP with:

--experimentalVision

Packaged releases include

  • MCP Bridge desktop app

  • Local MCP server

  • Chrome DevTools MCP

  • Electron/Node runtime

  • Full OpenAI tunnel-client

  • Production app/tray icons

You do not need to manually install Node.js, npm, Go, the MCP SDK, Chrome DevTools MCP, or tunnel-client when using an official packaged release.

Release artifacts

Platform

Artifact

Linux

MCP-Bridge-<version>-linux-amd64.deb

Linux

MCP-Bridge-<version>-linux-x86_64.AppImage

Windows

MCP-Bridge-<version>-win-x64.exe

Every installer also has a matching:

<installer>.sha256

The built-in updater refuses to install an update if its checksum is missing or does not match.

Build from source

Linux

git clone https://github.com/haseebgb92/ChatGPT-MCp.git
cd ChatGPT-MCp
npm install
npm start

Windows

git clone https://github.com/haseebgb92/ChatGPT-MCp.git
cd ChatGPT-MCp
npm install
npm start

A development run expects a full tunnel-client either on PATH or selected in Advanced settings.

Build installers

npm install
npm run check

Linux:

npm run dist:linux

Windows:

npm run dist:win

GitHub Actions builds the official release artifacts and SHA-256 sidecars from version tags.

Security

MCP Bridge deliberately exposes powerful local capabilities, so use the smallest access level that fits the task.

  • Never commit Runtime API keys.

  • Prefer a restricted Runtime API key.

  • Use separate tunnel IDs for Local and Web MCP.

  • Prefer read-only roots unless write access is required.

  • Mounted roots are real-path checked before use.

  • Enable shell only when required.

  • Full shell is not a sandbox.

  • Chrome MCP can act inside logged-in browser sessions.

  • Review consequential actions before allowing them.

  • Runtime logs redact OpenAI-style sk- keys.

  • Update installers are SHA-256 verified before installation.

See SECURITY.md for details.

Project structure

.
├── .github/workflows/build.yml
├── assets/
├── bundled/
│   └── local-mcp/
├── docs/
│   └── images/
├── install-linux.sh
├── install-windows.ps1
├── src/
│   ├── main.js
│   ├── preload.js
│   ├── updater.js
│   └── renderer/
├── LICENSE
├── package.json
└── README.md

Current release

v0.2.11

The latest public release includes the desktop updater, local files and Chrome connections, optional Meta Manager MCP integration, and privacy-reviewed support diagnostics. See all release notes for details.

Download the latest release

License

MIT License.

Credits

  • OpenAI Secure MCP Tunnel / tunnel-client

  • Model Context Protocol

  • Chrome DevTools MCP

  • Electron

Built by Advertpreneur.

Meta Manager MCP (v0.2.9)

  1. Install and open Meta Manager v0.1.21 or later on the same Windows or Linux computer. In Meta Manager → Settings verify that Local MCP is online.

  2. Create a separate tunnel ID using OpenAI Tunnels. Never reuse your Local or Web tunnel IDs.

  3. In MCP Bridge → Connection settings, enter the Meta Manager tunnel ID and your Runtime API key, then select Save settings.

  4. Choose Start Meta in Connections. The tunnel client uses native HTTP transport to http://127.0.0.1:38939/mcp over the secure OpenAI tunnel (no extra proxy process required).

  5. Connect the corresponding tunnel in ChatGPT. Both applications must be running to use local MCP tools; posts already scheduled through Meta Manager Cloud do not require either PC to remain on.

The Runtime API key is kept in secure local storage and is never included in installer files or public links. The Meta Manager local MCP service stays bound to 127.0.0.1 rather than a public HTTP listener. The tunnel ID alone is not a credential and cannot establish the connection without a Runtime API key and a running tunnel client.

Combined diagnostics and private GitHub report (v0.2.11)

Meta Manager v0.1.24+ records structured error categories for Facebook, Instagram, LinkedIn OAuth and MCP startup. MCP Bridge records its tunnel startup and runtime failures. Both applications keep their diagnostics in their own local app-data folders.

To send a reproducible support report:

  1. Reproduce the error in Meta Manager or MCP Bridge.

  2. In MCP Bridge, open Support diagnostics → Generate report.

  3. Inspect the preview, which includes only allowlisted error categories, optional HTTP status, timestamps, application version, OS and local MCP reachability.

  4. Choose Copy sanitized report, or Open GitHub issue draft.

  5. GitHub opens an issue draft in the private Meta Manager repository (GitHub access required). Only you can submit it.

The applications do not automatically post reports, do not include raw process logs, and do not include OAuth codes/states, account names, tokens, Runtime API keys, tunnel IDs, passwords, usernames, email addresses or full file paths. Keep raw traces on your computer and never paste them into public issues.

Workspace UI — v0.2.13

The desktop interface now uses seven isolated views instead of scrolling through a long dashboard: Overview, Connections, Local folders, Browser, Team peers, Activity and Settings. Each view has a dedicated header, focused controls, and its own scroll position. The sidebar collapses to icons at medium window sizes and becomes a horizontally scrollable tab bar in compact layouts.

All Local/Web/Meta Manager tunnel operations and Team Peers IPC remain unchanged. The header reports live tunnel status. Editing a setting or folder displays Save changes; Ctrl+S / Cmd+S also saves from any view. Keyboard users can move between tab headers with the arrow keys, Home/End, or use Alt+1 through Alt+7. Direct tab links use URL hashes (for example #peers). Tab transitions respect reduced-motion settings.

Verify responsive behavior and original renderer controls without connecting to real tunnels:

python3 tests/ui_layout_cdp.py

This isolated test uses local Chrome DevTools and a mocked Bridge interface. It checks seven views at 390, 700, 900, 1152 and 1440 px, guards against overflow and script errors, verifies keyboard navigation, dirty state, and Save integration. The test does not read production API keys or write production configuration. The Electron app uses the existing Local MCP code and current saved settings.

v0.2.14 Windows tunnel startup hotfix

v0.2.13 accidentally bundled Linux node and tunnel-client executables into the Windows NSIS installer. The Windows MCP service launcher requires resources/node-runtime/node.exe and resources/tunnel-client/tunnel-client.exe, so users without their own Windows copies could not start Local, Web, or Meta Manager MCP. This is a packaging regression, not a Tailscale or account/tunnel ID problem.

v0.2.14 packages platform-specific binaries and adds scripts/check-runtime-assets.js (pre-build) plus scripts/check-packaged-runtime.js (post-build) so a wrong-platform build fails before publication. The Windows binaries used for this hotfix were recovered from the official v0.2.11 Windows GitHub release, after verifying that release's SHA-256 checksum; the recovered node.exe SHA-256 is 9c9245166b4a8e182e0b797da9c20136117ff24368eaff1fec8343a123c8db0e and tunnel-client.exe is 839cb2cc21cdc4fd1ad793506e8cb98161518ac559ad88c984c9a0e993d89dad.

For future native Windows source builds, install-windows.ps1 bundles both locally installed node.exe and a freshly built tunnel-client.exe. For cross-compilation, first provide trusted Windows x64 binaries in vendor/node-runtime/node.exe and vendor/tunnel-client/tunnel-client.exe; the build refuses to proceed if they are missing or the wrong format. Linux packages include only Linux executables. The original configured tunnel IDs and encrypted API key remain in user data during in-place upgrades.

Team Peers: persistent approved control (v0.2.15)

For long-running tasks such as local AI model fine-tuning, control permissions now support 10 minutes, 1 hour, 8 hours, and Until manually revoked. Approval is specific to one paired computer. With Until manually revoked, approval survives MCP Bridge and system restarts; timed grants do not survive a restart. Screen sharing can separately be enabled and remains enabled across restarts until switched off.

The computer owner must explicitly select the trusted peer, choose the duration, and confirm the permission. A visible red control banner appears throughout the app while remote control is active; Revoke access now is available there and in the tray menu. Removing the authorized peer, regenerating the invitation token, or revoking permission disables remote command and input access. No control permission is granted automatically by updating.

The screen preview updates roughly once per second while the Team Peers tab is open; it is not full-motion video streaming. Remote shell commands still have an individual 20-second execution ceiling; launch long-lived training processes as persistent background jobs and monitor separate logs/checkpoints so a ChatGPT turn or the peer transport does not need to remain open. The owner's explicit revoke stops new remote commands/inputs but does not kill independently launched training processes.

Large remote screen preview (v0.2.16)

From Team Peers, select a computer and click Expand preview. The remote screen opens in a near-full-window viewer, begins screen sharing automatically if permitted, and keeps the existing mouse-click, remote text and command controls accessible. Choose Restore view, press Esc, or click the dimmed backdrop to return without disconnecting the peer. Clicking Stop viewing also closes the large preview.

Expanded preview uses a larger 1920×1080 JPEG screen capture, while compact preview and the existing peer_view_screen MCP tool continue using 1280×720 to reduce Tailnet traffic. Actual capture dimensions depend on the host display and Electron's capture implementation. The preview refreshes roughly once per second, not continuously like a full-video streaming client. Host consent to share the screen or permit remote control is unchanged.

The new preview and restore interactions passed UI automation at 1152×768 and 390×844, including immediate screen start, high-resolution request, keyboard Escape, correct modal sizing, and restoration to the original workspace. All existing MCP/peer unit checks and responsive layout checks passed.

Available Tools

5 tools
file_infoA
Read-onlyIdempotent

Inspect a file or directory in an approved root and return its type, size and modification time without reading its contents. Use read_file for UTF-8 content. Read-only.

Inputs: path (required), root (required).

list_filesA
Read-onlyIdempotent

List up to 1,500 immediate files and directories in an approved root, including relative path, type, size and modification time. Use search_files for recursive discovery. Does not change files.

Inputs: root (required), path.

list_rootsA
Read-onlyIdempotent

List the approved root names and access modes. Call this before other file tools to discover valid root values; filesystem paths and file contents are not returned.

Inputs: no named parameters.

read_fileA
Read-onlyIdempotent

Read text from an approved local root as UTF-8, truncating at MAX_READ_BYTES (512 KB by default) with a clear truncation marker. Use search_files first to locate passages; binary files are not supported. Read-only.

Inputs: path (required), root (required).

search_filesA
Read-onlyIdempotent

Recursively search filenames and UTF-8 contents inside an approved root with case-insensitive substring matching. Returns relative paths and brief snippets, bounded by max_results and a file scan limit; skips symlinks and build folders. Use read_file for complete text.

Inputs: root (required), query (required), path, max_results.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 5 tools

    Disambiguation4/5

    Each tool has a distinct role: discovering roots, listing directory contents, inspecting one path, reading text, and recursive searching. The only mild overlap is that list_files and file_info both return metadata, but their scope is clearly differentiated by the descriptions.

    Naming Consistency4/5

    Four of five tools follow a clear verb_noun convention (list_roots, list_files, read_file, search_files). file_info is noun-based rather than get_file_info, which is a minor deviation but still understandable.

    Tool Count5/5

    Five tools is well-scoped for a read-only filesystem bridge. Each tool covers a necessary operation—discovery, listing, metadata, reading, and searching—without redundant surface area.

    Completeness4/5

    The read-only filesystem surface is coherent and covers discovery, navigation, metadata, reading, and search. It lacks write/create/delete operations, but the descriptions consistently frame the server as read-only, so this is a minor scope gap rather than an internal inconsistency.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables ChatGPT Web to securely access local files and run commands via MCP, with optional OpenCode agent mode for autonomous tasks.
      -
    • F
      license
      Not graded
      quality
      C
      maintenance
      Enables ordinary ChatGPT web Chat to work directly on a local machine over a private MCP tunnel — reading and editing project files, running shell commands and tests, loading local skills, manipulating images, and driving Chrome — without spinning up a separate coding agent or additional model-generation API calls. It also supports multi-step task plans with persisted execution receipts and acceptance checks so results from real runs are returned to the conversation.
      -