- Home
- Skills
- DevOps & Cloud
- Medicine Man: Let AI Diagnose Proxmox Read-Only
More screenshots
Works with the AI Agents you already use
Medicine Man: Let AI Diagnose Proxmox Read-Only
Give your AI agent a read-only SSH key to your Proxmox hosts. A filter on each host refuses starts, stops, reboots, secret files and firewall changes. Your agent finds the fault and you decide the fix.
Free
Medicine Man: Let AI Diagnose Proxmox Read-Only
Example session with this skill installed
My Node app in CT 101 on my Proxmox box pve1 keeps going down at night. I'd like you to be able to look at the host and the container yourself when it happens, but I don't want you able to start, stop or reboot anything or open anything to the internet. It's a three node cluster. How do I set that up, what do you need from me, and how do I know it's really locked down?
- Read your context and instructions
- Compiled the medicine man
Use Medicine Man in gate-only mode. Each node gets an SSH forced command called healer-gate, and I get a read-only key that passes through it. The node refuses anything outside the allowed list, so the limit doesn't depend on me behaving. I can look at the host and inside CT 101. Any fix comes to you as an exact command for you to run.
No daemon runs in this setup, so I won't be watching at night. You ask me when it's down, or the next morning and I read the logs from the night. Automatic diagnosis needs the full daemon, but that holds a read-write key, so it goes further than you asked for.
What I need from you
- Your OK to
git clonethe code from https://github.com/AbdurahmanZA/medicine-man (MIT licence) onto the machine I run on. That machine needs bash 4 or newer. - The names and addresses of all three nodes. Below, pve2 and pve3 stand in for your other two.
- The app's service or pm2 name, its port, and roughly what time it goes down.
- You run the host steps with your own root access. Don't give me a root password or any other key.
On the machine I run on
git clone https://github.com/AbdurahmanZA/medicine-man.git ~/medicine-man
git -C ~/medicine-man checkout d6465b0
python3 ~/medicine-man/gate/test_gate.py # expect: failures: 0
ssh-keygen -t ed25519 -N '' -C healer-ro -f ~/.ssh/healer_ro
mkdir -p ~/.config/medicine-man
cat > ~/.config/medicine-man/hosts <<EOF
pve1 192.0.2.10
pve2 192.0.2.11
pve3 192.0.2.12
EOF
export HEALER_HOSTS_FILE=~/.config/medicine-man/hosts
export PATH="$HOME/medicine-man/bin:$PATH"
ssh-keyscan 192.0.2.10 192.0.2.11 192.0.2.12 >> ~/.ssh/known_hosts
Compare the scanned fingerprints with what each node shows for ssh-keygen -lf /etc/ssh/ssh_host_ed25519_key.pub.
On the nodes, as root
Keep a root session open on each node until the checks pass.
- Copy
gate/healer-gate.pyto all three nodes and run this on each. The key file is shared across the cluster, and a node without the script refuses the key.
install -m 755 healer-gate.py /usr/local/bin/healer-gate
- On one node only, append the key. Paste in the contents of
~/.ssh/healer_ro.pub. Use>>, because overwriting this file locks everyone out of all three nodes.
echo 'command="/usr/local/bin/healer-gate ro",restrict,from="AGENT_IP" ssh-ed25519 AAAA...your-public-key... healer-ro' >> /etc/pve/priv/authorized_keys
How to know it's locked down
Run these while CT 101 is up, so a failed denial does no harm. Repeat for pve2 and pve3.
python3 ~/medicine-man/gate/test_gate.py # failures: 0
px pve1 'pct list' # works
px pve1 'pct start 101' # DENIED
px pve1 'cat /etc/shadow' # DENIED
px pve1 'pct list; pct start 101' # DENIED
ssh -o BatchMode=yes root@192.0.2.10 id # must NOT print uid=0
The last line matters most. The gate only binds this one key. If my account already has another root key or an ssh-agent that gets into these nodes, I have unfiltered root and none of this limits me. Remove that access or run me from a different account.
Read the log. /var/log/healer-gate.log on each node records every allow and deny, so you can check what I actually tried, then and later.
- If a denial doesn't appear, delete the
healer-roline from the key file and stop.
Don't use reboot or pct stop as tests. If the gate were broken, they would really happen.
What it doesn't protect
- Everything I read goes to the model provider. Known secret files are blocked, but log lines and service files can still contain personal data or secrets.
pct execruns as root inside CT 101. The gate checks the command string and can't see what a program inside the container does, so the container is less tightly held than the host.- The private key has no passphrase, so protect that file. Anyone who gets it has the same read-only access.
- Read-only commands can still be slow on a busy host.
- The test suite covers known attacks. It can't prove there is no bypass, so "locked down" here means tested and logged, not guaranteed.
When it goes down
I'll check whether CT 101 is running, then the host's disk, memory and pool health. Inside the container I'll read the failed units, the app's journal or pm2 logs, the listening ports and a local request to the app.
If CT 101 is stopped, I'll report that and won't suggest starting it, since only you know whether it was shut down on purpose. If I get a DENIED, I'll treat it as final and tell you rather than try another way round. You'll get the evidence, the likely cause, and any fix as a single command to run on the node.
Connects securely to your tools. The creator never sees your data.
About this skill
You want your agent to look at a broken Proxmox server. You don't want it starting a container you shut down on purpose, or opening a port.
Medicine Man puts a filter on each Proxmox host as an SSH forced command. The agent gets a read-only key, and the host decides what that key may run.
Your agent will
- Install the gate on one host or a whole cluster, and prove it refuses what it should
- Diagnose a down service step by step: guest, host, container, app, then the path to it
- Report a stopped guest and leave it stopped
- Give you each fix as one exact command to run yourself
- Write playbooks that limit which fixes may run without you
The code is open source under the MIT licence. This skill fetches it from GitHub with git clone, at a fixed commit.
How it was tested
The gate has a test suite of 212 allow and deny cases, which runs on every change. An independent review found ways past an earlier version, and those are fixed and added as tests. Tests can't prove there is no bypass, so read the limits below.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for DevOps & Cloud.
Inside a VM the agent sees only status and config. Full diagnosis needs an LXC container.
The gate limits one key. An agent that holds other root access is not limited by it.
Log lines and service files can still contain personal data or secrets.
Tests can't prove there is no bypass.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 8 days ago
- Free with an Agensi account
- Software
- Python 3 and gitbash 4 or newer on the agent's machine
Git
A Proxmox VE host with root access for setup.
Permissions
File Scopes
Fetches the open-source gate from GitHub with git clone, at a fixed commit. Uses SSH to reach only the Proxmox hosts you list. Everything the agent reads goes to your model provider.