AI chat and act mode

Ask a question

Open Ask AI in the top navigation. Pick a server from the Pin a server for this chat menu, or leave it on All servers for fleet-wide questions. Once a conversation has started, its server context is fixed; start a new conversation to change it.

Starter questions show the kind of thing that works well: "Why is web-01 slow?", "Which servers have SSL expiring soon?", "Show me all critical issues across my infrastructure", "What changed in the last 24 hours?". You can also paste an alert or monitoring email; the AI treats it as a question about what happened at that time and checks the incident timeline for that window.

Answers stream in as they are written. Conversations are saved and listed in the sidebar so you can come back to them.

What the AI can read

The AI does not browse your servers. It calls a fixed set of read-only tools, each limited to your workspace. In plain words they are:

• •get_server_overview — CPU, load, memory, disk, uptime and the top processes by CPU on one server
• •get_recommendations — the current ranked findings for one server, optionally by severity
• •get_log_summary — the latest plain-English log summary and top log clusters from the last hour
• •get_incident_timeline — correlated events and the likely root cause over the last N hours (up to 30 days)
• •get_services — every detected systemd, Docker and PM2 service and its status
• •get_domains — the domains a server serves and their SSL expiry
• •get_kubernetes_snapshot — nodes, workloads, problem pods and recent events for a connected cluster
• •list_agents — every server in the workspace with its health status
• •get_recent_commands — commands actually run on a server recently, with exit codes
• •propose_fix — presents a ready-to-run fix as a card with risk and reversibility; nothing runs
• •ask_choice — asks you a short multiple-choice question as clickable buttons

A normal turn may use up to 5 tool calls. Under each answer, Data sources used lists exactly which tools were called, so you can see what the answer is based on.

Act mode: let the AI fix things

By default the chat only proposes fixes. Tick Let AI fix things in the composer and the AI gains one extra tool, run_command, for that conversation. It can then diagnose, run the fix, and run a check afterwards to confirm it held. The tool budget rises to 12 calls per turn.

What has to be true before a command runs:

• •You are a workspace owner or admin. Other roles do not see the toggle and get a clear refusal if they try.
• •If your workspace has Require a 2FA code before server actions turned on, you confirm your identity first.
• •The target server has remote actions switched on, is online, and the workspace is on the Growth plan or higher.

Each command goes through the same pipeline as a command you type in the terminal: it is Ed25519-signed, recorded in the server’s command history, and shown in the chat as an AI ran a command card. The card shows the command, then "Running on the server…", then the result: Finished · exit 0, a non-zero exit code, Not run if a gate refused it, or Agent did not pick it up if the server never claimed it. The cards are kept with the conversation.

The AI runs read-only checks freely and runs ordinary fixes you asked for (restart a failed service, renew a certificate, clear a full log) without asking again. For anything destructive or hard to reverse, such as deleting files, rebooting, stopping a database or changing firewall or SSH access, it stops and asks first with ask_choice.

Credits and running out

Every AI call costs credits. Your plan grants a monthly allowance; the balance is shown in the top bar. When the balance reaches zero, any AI action returns an Out of AI credits prompt with three choices: top up, connect your own key, or wait for the allowance to reset at the start of next month. Monitoring, alerts and the rest of the dashboard keep working without credits.

Bring your own key

Under Account → Developer → Bring your own AI key (BYOK) an owner or admin can connect a provider key for the whole workspace. While it is active, AI runs on your account and platform credits are paused. Supported providers:

• •Anthropic (Claude)
• •AWS Bedrock (Claude) — access key, secret and region
• •OpenAI
• •Azure OpenAI — endpoint and deployment name
• •Google Gemini
• •OpenRouter

You can name a model for the provider; otherwise a sensible default is used. The key is stored encrypted and only its masked form is ever shown again.

Ask from Slack

Under Account → Notifications → Slack ChatOps, click Add to Slack. That installs Opservo into your Slack workspace and sets up the /opservo command. Type /opservo why is web-01 slow? in any channel; Slack shows "On it — reading your servers…" and the grounded answer follows in the channel.