Proxmox server controlled from chat by a Gemini agent

Proxmox server controlled from chat by a Gemini agent

An n8n workflow with a chat window. You describe what you want done on a Proxmox cluster, such as listing the VMs on a node
or starting one. A Gemini agent turns the request into a Proxmox API call with a method, a path and a body. The workflow
runs the call with your API token, and a second agent explains the result in plain words.

What is inside

• workflow.json: the workflow as n8n exports it, 24 nodes beside 7 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

Mindket cleaned up and repaired the copy:

• Three trigger nodes led nowhere: Telegram, Gmail and a webhook. Each would have asked for a credential before the
  workflow could be published, so Mindket removed them. A sticky note still says any trigger can feed the agent.
• A sticky note held a donation link and personal contact details. Mindket's copy drops it.
• The original copy held someone's own Proxmox addresses: one public host and some private ones. They are now
  <YOUR_PROXMOX_HOST>, <YOUR_PROXMOX_HOST_2> and <YOUR_PROXMOX_HOST_3>, and one request node that called a different
  host from the rest now calls the same one.
• The GET branch read its path from output.properties.url.pattern, a field the agent's answer does not have. The prompt
  and the output example put it in url, as the other branches read it, so Mindket's copy reads output.url there and in
  the summary agent's text.
• The three chat models named models/gemini-2.0-flash-exp. Google's deprecations page, read on 12 November 2026, says the
  gemini-2.0-flash family shut down on 1 June 2026 and names gemini-3.6-flash as its replacement. The page does not list
  the exp name. Mindket's copy uses models/gemini-3.6-flash.
• No credential is inside.

Before you start

• An n8n instance, on n8n Cloud or your own server, that can reach your Proxmox on port 8006.
• A Proxmox API token. A token that can only read is the safe way to start.
• A Google Gemini API key, for a Google Gemini (PaLM) API credential.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Proxmox, create an API token. In n8n, add a Header Auth credential with the name Authorization and the value
   PVEAPIToken=user@pam!tokenname=<token-secret>, using your own user, login type and token name.
3. Select that credential in Proxmox, HTTP Request, HTTP Request1, HTTP Request2, HTTP Request3 and
   HTTP Request4.
4. In those same six nodes, replace <YOUR_PROXMOX_HOST> with the address of one of your nodes.
5. In the prompt of AI Agent and the description of Proxmox, replace the three host placeholders with your own addresses
   and psb1, psb2 and psb3 with your node names. Delete lines for nodes you do not have.
6. Add a Gemini credential and select it in Google Gemini Chat Model, Google Gemini Chat Model1 and
   Google Gemini Chat Model2.
7. Save, open the chat and ask for something harmless, such as listing the VMs on a node.

How a run goes

1. When chat message received starts a run when you send a message.
2. AI Agent is a ReAct agent with Google Gemini Chat Model. It can call three tools: Proxmox, which reads the cluster
   status, and Proxmox API Documentation and Proxmox API Wiki, which read two pages from Proxmox's site. It answers in
   JSON with response_type, url and details. Structured Output Parser sets that shape, and Auto-fixing Output Parser
   asks Google Gemini Chat Model1 to repair an answer that does not fit.
3. Switch sends the request by response_type: GET, POST, PUT, OPTIONS, DELETE or Invalid.
4. For GET, HTTP Request calls the path. Structure Response turns each item into one line of text, and AI Agent1, a
   Conversational agent with Google Gemini Chat Model2, explains it in words.
5. For POST, If checks whether the answer has details. With them, HTTP Request1 posts them as the body. Without,
   HTTP Request2 posts with no body. Merge joins the two.
6. For DELETE, If1 leads to HTTP Request3 or HTTP Request4. Both send DELETE with no body, and Merge1 joins them.
7. Structgure Response from Proxmox splits the task id that Proxmox returns into node, operation, user and time. `Format
   Response and Hide Sensitive Data` writes one sentence from them.

Change the look

• What the agent may do and how it writes calls: the prompt in AI Agent.
• Taking deletion away: remove the DELETE rule in Switch, or the nodes after it.
• The input: swap the chat trigger for another trigger. The prompt reads {{ $json.chatInput }}.
• The summary's tone: the prompt in AI Agent1.
• The models: the three Gemini nodes.

Good to know

• It changes things. POST and DELETE calls run at once, with your token's rights, from what a model understood. There is no
  confirmation step, so a request to delete a VM deletes it. Use a narrow token and a lab you can rebuild.
• The five HTTP Request nodes have the option that ignores SSL certificate problems turned on, so they trust any certificate. That
  suits a Proxmox with its own certificate, but turn it off if yours is valid.
• PUT, OPTIONS and Invalid have nothing wired to them, so the run ends with no call and no answer. The agent's own examples
  include PUT, for things such as resizing a disk, so those requests are lost.
• n8n marks the ReAct and Conversational agent types used here as deprecated and says they are removed in version 3. A newer
  AI Agent node replaces them.
• Your requests and what Proxmox returns, such as VM names and status, go to Google's Gemini.
• It runs from n8n's Chat window in the editor. Nothing in the file opens a public chat.
• The JSON holds no credential.