Install and prepare the machine
The desktop application manages the operational layer, but the local MCP runtime it starts still needs to exist on the machine.
Install Tunnel ManagerUse the signed Windows installer once the public stable release is available. The landing page and release endpoint will expose the same version metadata.
Prepare the MCP runtimeInstall the runtime required by your target MCP server. For Serena-based projects, make sure the Serena command can be resolved from the environment used by Tunnel Manager.
Keep the project localYour source project remains on the machine. Tunnel Manager only coordinates the local MCP process and secure tunnel connection.
Create a profile
A profile is the reusable connection definition for one MCP workspace. Keep one profile per project/runtime combination so validation and runtime state remain unambiguous.
Recommended profile fields
Project folderThe local workspace that the MCP server should operate against.
MCP target commandThe command Tunnel Manager starts, for example a Serena MCP server for the selected project.
Tunnel identity and runtime keyUse the values issued for the OpenAI secure tunnel workflow. Sensitive values should be stored only in the local control layer.
Validate before start
Validation should be treated as a required preflight step. A healthy profile must pass configuration, credential, tunnel identity, MCP target, and local runtime checks before the Start action becomes available.
Connect ChatGPT
Once validation passes, start the profile. Tunnel Manager starts the configured local MCP process, establishes the secure tunnel, and exposes the connection state in the desktop UI.
Start the validated profileWait until both the MCP runtime and tunnel show a healthy running state.
Open ChatGPT connector settingsSelect or reconnect the MCP connector associated with the tunnel.
Watch AI activityWhen ChatGPT invokes a local MCP tool, Tunnel Manager should transition from Connected/Standby to AI Active and stream the corresponding runtime logs.
Serena project setup
For Serena-backed projects, keep project-level behavior in the project’s .serena directory and machine-specific overrides in project.local.yml.
Trusted project requirement
Serena only runs a project activation_command when the project path is trusted by the global Serena configuration. Prefer trusting the exact project path or a narrow development directory instead of using a global ** rule.
Troubleshooting
Start with the visible validation state, then inspect the streaming logs. The problem is usually isolated to one layer of the chain.
Confirm the profile loads the existing stored key before adding another one. Revalidate after any credential change.
Verify the tunnel belongs to the intended environment and that the profile is not reusing an obsolete identifier.
Confirm the project folder exists on the machine where the MCP runtime actually runs. Remote Linux workspaces must use a path valid on that Linux host.
Run the configured MCP target manually in the same user environment. Fix PATH/runtime installation before retrying Tunnel Manager.
Add the exact project path to Serena’s global trusted_project_path_patterns, then restart the Serena/tunnel process so the new global configuration is loaded.
Identify the existing process first. Avoid starting duplicate servers; either reuse the healthy process or stop the stale process and restart the profile.
Check local network reachability, profile credentials, and tunnel state. If auto recovery is enabled, verify whether the reconnect attempt is progressing in the live log.
Check the current state. Start should be disabled while a process is already starting or running; Validate may also be unavailable during incompatible transitions.