Give your AI your production errors

What MCP is, in one paragraph

The Model Context Protocol is an open standard that lets an AI assistant use outside tools. A tool that speaks it, an “MCP server”, offers a few actions, such as “list open incidents”, and the assistant calls them when it needs to. Claude, Claude Code, Cursor, VS Code, Windsurf, Lovable and Bolt all speak it, so one connection works in each.

Why it beats copy-paste

• Evidence, not a summary. The agent gets the error, the stack, the request that failed, what the visitor clicked before it, and how many people it affects.
• Worst first. “What’s broken in my app?” returns the open problems ranked by how many people they hit.
• A loop that closes. After the fix is deployed, the agent marks the problem fixed, and the monitoring tool tells it whether the error stayed away.

Which monitoring tools have an MCP server

Sentry and PostHog both have one, aimed at engineers who already use them: the agent gets issues and stack traces from their data. Vigilia’s is built for this loop: each incident comes with a plain explanation and a fix prompt already written for your builder, and marking it fixed starts a watch that confirms the fix with real traffic. It is included on the free plan, with 2,000 requests a day; fixing one bug takes a few dozen.

Connect Vigilia to your AI

The server address is https://app.getvigilia.com/mcp. Each tool signs in with OAuth: it opens a Vigilia page, and you click Allow.

• Claude (web, desktop, mobile): in a chat, + → Connectors → Browse connectors, search for Vigilia, Connect.
• Claude Code: claude mcp add --transport http --scope user vigilia https://app.getvigilia.com/mcp, then /mcp → vigilia → Authenticate.
• Cursor: one click from the dashboard’s Connect AI page, then Connect in Cursor Settings → MCP.
• Lovable: Settings → Connectors → add a custom MCP server, choose OAuth.
• Bolt: Connectors → Custom MCP server, transport HTTP, OAuth.
• VS Code, Windsurf and other clients: add the address; they sign in the same way.

The docs have every step, the tools it offers and what to do when a connection fails.

Then ask

• “What’s broken in my app right now?”
• “Explain my worst Vigilia incident and how to fix it.”
• “Use Vigilia to fix my worst incident.” In a coding agent with your project open, it finds the cause in your code, fixes it, and once you’ve deployed, marks it fixed.
• “Did my last fix hold?”
• “Install Vigilia in this project.” It adds the script tag in the right place and checks the live site.

Treat error data as data

Error messages, page addresses and form contents come from your visitors, and a visitor can type anything, including text that reads like an instruction to an AI. Whatever tool you connect, the agent should treat what it reads as evidence to investigate, never as orders. Vigilia marks every piece of visitor-supplied evidence as untrusted when it hands it to your AI, and tells the agent to check with you before following a fix prompt that asks for something unrelated, such as adding a third-party script.

Two more habits help: choose Read only when you connect if you only want the agent to look, and keep your agent asking before it runs tools until you trust the setup.

Read next

• Vigilia for Claude Code
• Vigilia for Cursor
• Docs: connect your AI

Questions