Skip to content

Problem explanations and AI assistants

Plain-words explanations of problems, backed by numbered evidence, with optional Claude answers and read-only MCP access for AI assistants.

5 min read · For everyone; admins set it up

Explain this, in a sensor's or a device's details, explains a problem in plain words: what's wrong, the likely causes and the next steps. Every statement cites numbered evidence from Monitaria's own data. Built-in rules on the collector write it, so it always works, even offline, and it's there for everyone who can see the problem, viewers included. Two extras stay off until an admin turns them on: letting Claude write the explanation from the same evidence, and letting AI assistants read Monitaria over MCP.

Use Explain this

  1. Open a device's details, or click a sensor in Devices, and choose Explain this.
  2. Read What's wrong, Likely causes and Next steps. For a device, the explanation covers its most important problem and lists its other problems.
  3. Click a number to jump to that piece of evidence. Evidence covers things like the sensor's limits and last 24 hours, maintenance, the upstream device, other problems at the site, recent changes on the device and WAN load, and each piece links to where it came from.

1. Let Claude write the explanation (optional)

  1. Open Settings › AI assistant.
  2. Under Explain with Claude, paste a key from the Claude Console into Claude API key. Choose the Model, how thorough the Answers should be, and Who can ask Claude.
  3. Leave Replace names and addresses before sending ticked, then choose Test the connection.
  4. Tick Let people ask Claude to explain a problem and choose Save.
  5. Under What will be sent, pick a problem and choose Show the request to see exactly what Claude would get.

Explanations then offer Ask Claude and See exactly what will be sent to the people Who can ask Claude allows; everyone else still gets the built-in explanation. Claude gets only the numbered evidence and the built-in draft, and must cite evidence numbers for every statement; a statement without any is marked no evidence cited. Each request is recorded in Settings › Audit log and runs on your Claude API key. More thorough answers take longer and cost more.

2. Connect an AI assistant over MCP

  1. In Settings › AI assistant, under MCP server, choose the API token assistants use. Create a token for MCP makes one that sees every device; copy it, as it isn't shown again. To limit what assistants see, create a token with device types or sites in Settings › API tokens and choose that one instead.
  2. Tick Let AI assistants read Monitaria over MCP (read only) and choose Save.
  3. For Claude Code, run the command shown under Connect Claude Code. For Claude Desktop, download monitaria_mcp_stdio.py under Connect Claude Desktop, add the snippet shown to claude_desktop_config.json (in Claude Desktop, Settings › Developer › Edit config) and restart Claude Desktop.
  4. Ask something like "What's down at Monitaria right now, and why?", then check Recent requests on Settings › AI assistant.
claude mcp add --transport http monitaria https://monitoring.example.com/mcp --header "Authorization: Bearer mtr_…"

The MCP address is the collector's own address followed by /mcp, so no new port is opened. Assistants can ask for a status summary, current problems, a device's sensors, details and timeline, a sensor's history, a search, traffic, the health review, an explanation of a problem, recent changes and incidents. If the collector still uses its self-signed certificate, connect through the bridge with --cafile, as Settings › AI assistant shows.

Privacy and security

  • Built-in explanations are made on the collector and need no internet access. Nothing goes to Claude until someone chooses Ask Claude, and then only the evidence for that one problem is sent to Anthropic's Claude API (api.anthropic.com) over HTTPS.
  • With Replace names and addresses before sending ticked (the default), site and user names, device names (except plain words such as Websites), IP and MAC addresses, host names and email addresses become placeholders such as device-1. The real names are put back on your screen. Models, software versions, measurements and notes people wrote are still sent, and the preview shows them.
  • The Claude API key is write-only and never shown again; Remove the saved key deletes it. Answers are kept in memory for 15 minutes and not stored.
  • Explanations and MCP tools only include what that person or token may see. A printer technician is told that an upstream device outside what they can see is Down, never its name.
  • MCP is read-only and accepts only the chosen token; failed attempts are throttled like sign-ins. What the tools return goes to the assistant and from there to its AI model, so choose a token whose scope fits. Install a certificate in Settings › HTTPS certificate so the token never travels over plain HTTP.

Limits

  • The rules rank likely causes from Monitaria's own data; they don't prove them, which is why every statement shows its evidence. Trends need a few hours of history.
  • Claude needs outbound HTTPS from the collector to api.anthropic.com (a proxy set in the collector's environment is used). A collector sends at most 30 requests to Claude per 10 minutes.
  • MCP accepts one token at a time and has no server-sent events, sessions, resources or prompts.