Less Dashboard, More Agent: Day-2 Ops with the Elestio MCP
I have a confession: I used to keep the Elestio dashboard pinned in a browser tab all day. Deploy something here, resize a database there, hunt through logs when a service started throwing 502s. It works. It is also a lot of clicking for things I could describe in one sentence. So when we shipped the Elestio MCP connector, I did what any reasonable person would do and tried to stop opening the dashboard entirely. Here is what that actually looks like, and where I still keep my hands on the wheel.
We already wrote the setup walkthrough, so I am not going to re-explain connecting it. Thirty-second version: add https://mcp.elest.io as a custom connector in Claude, ChatGPT, Cursor, or Claude Code, authorize once through your browser, and you are in. This piece is about the part that comes after, the day-2 stuff.
What the connector actually exposes
The Elestio MCP maps 68 tools straight onto the Elestio API. That is the whole platform, not a toy subset: deploy and manage services across 400+ templates, handle volumes, firewall rules, SSL certificates, and SSH keys, run backups, snapshots, CI/CD pipelines, and system updates, pull container logs, and read billing and audit data. Anything you can do by clicking, an agent can do by asking.
The reason I trust it near production is the auth model. It uses OAuth 2.1 with PKCE: you authorize once in the browser, and the AI client gets short-lived tokens instead of a permanent API key sitting in a config file. Your credentials never leave the Elestio domain. That detail is the difference between "fun demo" and "thing I will actually point at a live service."
The workflows that replaced my dashboard
Once it is connected, the value is not "deploy WordPress by chatting." It is the boring operational work that normally costs you a context switch. A few I lean on:
| What I type | What the agent does |
|---|---|
| "Why is my Ghost service throwing 502s?" | Pulls recent container logs, reads the stack trace, points at the cause |
| "Resize the Postgres node to 8 GB before tonight's launch" | Resizes the service and confirms the new spec |
| "Snapshot this service before I change the env vars" | Triggers a backup, then makes the change |
| "Restart the worker and lock it so nobody touches it" | Restarts, then locks the service against changes |
| "What did I spend across all services this month?" | Reads billing data and totals it up |
The debugging one is the keeper. Instead of opening the dashboard, finding the service, opening the log view, and scrolling, I ask a question and get an answer with the relevant lines already quoted. The agent has the logs and the context, so it can reason about the error instead of just showing it to me.
The other quietly useful one is the "snapshot before I do something dumb" pattern. I am far more willing to experiment on a live service when taking a backup is a clause in a sentence rather than a five-click detour I will skip when I am in a hurry.
Where I still keep my hands on the wheel
This is the part the breathless agent demos skip. Letting a model operate real infrastructure is great until it confidently does the wrong thing at scale. So a few rules I follow:
- Destructive verbs get confirmation. Delete, move, and resize all change or risk data. I want the agent to tell me what it is about to do and wait, not surprise me. Treat anything irreversible as a human checkpoint.
- The audit trail is your friend. Every action the agent takes through the MCP shows up in your Elestio audit log, the same as any other API call. After an agent-driven session, I skim it. It is also how you answer "wait, what changed?" three days later.
- Scope the session to the task. Short-lived tokens already limit the blast radius, but it helps to keep an agent session focused on one job rather than handing it the keys and walking away.
None of this is exotic. It is the same discipline you would want before giving a new teammate production access. The MCP just makes the access conversational, so the guardrails have to be conversational too.
Troubleshooting
- Connector shows red or won't authorize. Re-run the browser authorization; the token may have expired. Confirm the URL is exactly
https://mcp.elest.iowith no trailing path. - The agent can't see a service. Make sure you authorized with the account that owns it. The MCP only exposes what your Elestio account can see.
- It hesitates on a destructive action. That is the design, not a bug. Confirm explicitly, or do that specific step in the dashboard if you would rather click it yourself.
So, is the dashboard dead?
Not quite, and I would be lying if I said I never open it. But the balance has flipped. Routine ops, debugging, scaling, backups, cost checks, now start as a sentence, and I reach for the dashboard mostly when I want to see something visually. That is the shift the whole industry is feeling right now: less clicking, more describing.
The connector itself is free with any Elestio account; you only pay for the services you actually deploy. If you have not wired it up yet, point your AI client at mcp.elest.io and start with something low-stakes like "list my services." You will probably keep going.
Thanks for reading ❤️ See you in the next one 👋