Personal AIVA MCP Bootstrap

A clean, self-hosted MCP server for a fresh Oracle Cloud Ubuntu VM. It gives ChatGPT controlled access to the user's own server through tools for shell commands, files, and personal skills.

Give this repository to an agent

Tell the agent:

Clone this repository on my Oracle Ubuntu server and follow AGENTS.md exactly. Do not stop until ChatGPT has scanned the MCP tools successfully.

Manual install

git clone https://github.com/banddude/aiva-bootstrap.git
cd aiva-bootstrap
sudo ./install.sh
sudo tailscale up
sudo tailscale funnel --bg 8000
tailscale funnel status

Take the HTTPS Funnel address and append /mcp.

Example:

https://oracle-name.example-tailnet.ts.net/mcp

Connect to ChatGPT

Current ChatGPT setup:

1. Use ChatGPT on the web.
2. Open Settings → Apps → Advanced Settings and enable Developer mode.
3. Open Settings → Apps → Create.
4. Enter the Tailscale Funnel /mcp endpoint.
5. Select No authentication. The URL is an unlisted, user-controlled endpoint; do not share it.
6. Select Scan tools, then create the app.
7. Test by asking ChatGPT to call health.

Full write-capable MCP support may depend on the user's ChatGPT plan and workspace permissions. Pro accounts can have more limited custom-app capabilities than Business or Enterprise/Edu accounts.

Installed tools

• health
• run_command
• list_files
• read_file
• write_file
• get_file_base64
• list_skills
• get_skill
• save_skill

Security model

• The MCP process runs as a dedicated aiva Linux user.
• File tools are restricted to /opt/aiva/workspace.
• Skills are restricted to /opt/aiva/skills.
• The service has systemd hardening.
• The Oracle firewall does not need port 8000 exposed.
• HTTPS exposure is provided by Tailscale Funnel.

The run_command tool intentionally provides broad command execution inside the AIVA workspace. Only connect this server to a ChatGPT account you trust.

Useful commands

sudo systemctl status aiva-mcp
sudo journalctl -u aiva-mcp -f
sudo systemctl restart aiva-mcp
tailscale funnel status