MCP · AI AGENTS

Don't give an AI root on your clients' servers.
Give it Cloudnan.

Claude Code, Codex, Hermes and OpenClaw connect through Cloudnan's MCP server. They get real server tools; you keep previews, approvals, an audit log and an undo for everything they do.

cloudnan.com / approvals

Approval needed

Claude Code wants to change web-01

Restart php8.3-fpm and reload nginx

Undoable

- pm.max_children = 5

+ pm.max_children = 12

/etc/php/8.3/fpm/pool.d/shop.conf

Health check after apply: https://shop.example.com returns 200. Rolls back automatically if it fails.

How it works

What happens under the hood.

  1. 01

    Create a scoped MCP token

    Pick which servers and which tool scopes the token allows. It is revocable at any time and never carries an SSH key.

  2. 02

    Add Cloudnan to your agent

    Paste the MCP endpoint into Claude Code, Codex, Hermes or OpenClaw. The agent discovers Cloudnan tools automatically.

  3. 03

    Every call is previewed

    Before a tool changes the server, Cloudnan computes the plan and labels it Undoable, Partial or Not undoable.

  4. 04

    Risky calls wait for you

    Fully undoable calls can run straight away; Partial and Not undoable calls wait for a human approval. The policy is yours to tighten.

  5. 05

    Health-checked and audited

    Applied changes are health-checked and rolled back automatically on failure. Every call lands in the audit log with the agent name.

  6. 06

    Undo in one click

    If the agent got it wrong, undo the change from the Changes timeline. No reverse-engineering what it typed in a shell.

What you get

The capabilities that ship.

Works with your agent

Claude Code, Codex, Hermes and OpenClaw today; any MCP-capable client can connect.

No raw root shell

Agents act through typed tools (edit config, deploy site, run migration) that Cloudnan can preview and reverse, instead of arbitrary commands.

Per-client scoping

Limit a token to one client's servers. Revoke it and the agent loses access immediately, with no SSH keys to rotate.

Approvals where you are

Pending actions wait in the dashboard with their preview and label, and you are notified by email, Telegram or WhatsApp.

Readable audit log

Each entry shows which agent acted, the plan, the result and any undo.

Same tools as the dashboard

Deploys, databases, backups, firewall, logs and the AI Server Doctor are all exposed as MCP tools.

FAQ

Common questions.

Why not just give Claude Code SSH access?
A raw shell gives the agent everything and leaves you with nothing to review or reverse: commands are hard to audit and there is no undo. Through Cloudnan MCP the agent uses tools that are previewed, approved when risky, logged, and undoable, and the token can be revoked any time.
Which agents are supported?
Claude Code, OpenAI Codex, Hermes and OpenClaw are tested. Any client that speaks the Model Context Protocol can connect with a Cloudnan MCP token.
What if the agent does something that cannot be undone?
Cloudnan labels it Not undoable in the preview, before it runs, and risky actions wait for your approval. You decide with the label in front of you.
Does this cost extra?
MCP access is part of the Cloudnan subscription. Your AI agent subscription (Anthropic, OpenAI, etc.) is billed by that vendor.

Related capabilities

Undo

How every change becomes reversible.

Read

AI Operations Assistant

Cloudnan's own built-in assistant.

Read

Drift Detection

Know when something changed outside Cloudnan.

Read

Put your servers on autopilot.

Connect your cloud account or bring any Linux server. Every change gets a preview and an undo, including the ones your AI agent makes.

Connect your cloud Bring your own server
  • 7-day free trial
  • No credit card
  • Your provider bills your servers
  • Cancel anytime

Running many client sites? Talk to us