AI agents

Use the AI you already have. Nothing is wired to one vendor, and every command an AI wants to run pops up for you to approve first: the exact command, the server and the AI's reason.

Claude Code (and any MCP app)

gridcity mcp

It prints a claude mcp add … command (and JSON for other MCP apps such as Cursor or Claude Desktop). Run it once; the app must be running when the AI calls it. In the app, Ctrl+K → Ask an AI agent… shows the same command.

Then ask in plain words:

• "Why is the site down?"
• "Which server is running out of disk, and what is filling it?"
• "Update the packages on the staging servers."

The tools

Tool	What the AI gets
incident_report	everything not healthy, grouped by likely root cause, and what is at risk
security_findings	attacks and exposures, and the checks that could not run
diagnose	the Doctor for one server or resource, with proven fixes
get_topology	links, and the blast radius of a building
list_servers	servers with status, CPU, memory, disks and discovered services
list_resources	databases, buckets, probes and load balancers with their metrics
list_runbooks	saved commands and where they apply
run_command	runs a shell command on one server after you approve it
wait_for_command	collects a command that was still waiting for you

Reference: MCP tools.

Approval cards

Cards appear on top of everything, even an open terminal:

• Allow, Allow all on this server (until the app restarts), or Deny.
• Add a note for the AI: instructions or why you deny ("use apt, not snap").
• When a command needs a secret on stdin (a sudo password, a token for --password-stdin), the card asks you for it. It goes to the command only, never to the AI, never to the log.

Allow all on this server

Allow all lets the AI work on that server without a card for every command, until the app restarts or you untrust the server in the agent panel. It is never saved.

• Everyday commands run straight away: status checks, logs, restarts, installs, upgrades.
• Destructive commands still ask, and the card says why: deleting files, wiping disks, dropping database data, removing containers or packages, deleting users or cron jobs, stopping services, killing processes, clearing firewall rules, discarding git history, deleting cloud resources, rebooting or shutting down.
• The AI can ask too: it marks any command it judges hard to undo as risky, and that command asks even on a trusted server.
• The sudo password is asked once. Type it on the card where you choose Allow all, and the app keeps it in memory for that server and reuses it for later sudo commands. It is never written to disk, logged or shown to the AI. If sudo rejects it, the app forgets it and asks again.
• Secrets the AI asks for, such as a registry token, are asked every time.

Prefer Allow on servers that matter: command output can contain text that tries to steer the AI.

An AI is never left hanging: if you do not answer within 25 seconds, it is told to wait and collects the result later with wait_for_command. Unanswered requests are denied after 5 minutes. Every command run or denied is in the audit log, with who allowed it.

Codex and Gemini CLI inside the app

Ctrl+K → Ask an AI agent…, pick an agent and Start. The app runs the vendor's open-source adapter through the Agent Client Protocol and you sign in with your own account (ChatGPT, Google, or an API key). You can type or talk to it. It works on servers only through the same tools and approvals.

Rules the AI follows

The tools tell the AI to explain why in each command, to prefer read-only checks before changing anything, never to ask for passwords in chat, and to apply Doctor fixes only with your click in the app.