Your assistant gets the verdict, not the rows.

Signal9 runs an MCP server, the open standard your AI assistant already speaks. Most MCP servers wrap a REST API and leave the model to work things out from raw records. Signal9 tools decide first, then hand over the decision and the evidence behind it, stripped to what an outside reader should see.

What crosses the wire. The live read, what is happening right now, open alerts, storm state, open incidents, who is being paged. The Rundown, what needs you, what handled itself, and what cleared last month but came back last night. Per-alert verdicts that carry their evidence and their windows, with honest abstention when history is thin. Records in full, with journal notes written as you and stamped as assistant-delivered. Current on-call coverage. Operational memory capture in your own words. Webhook setup with your account's real ingest URLs and a test loop through the live pipeline. And feedback, sent only after you approve the text.

Keys are scoped, a read-only key cannot see the write tools at all, and anything that writes acts as you, under your permissions. Connect in three steps, mint an Assistant, full access API key under Profile, then API Keys, point your MCP client at signal9.app/api/mcp over streamable HTTP, and send the key as a Bearer token.

What is the Signal9 MCP server? An MCP endpoint built into Signal9 that lets your AI assistant see what is happening right now (open alerts, storm state, who is being paged), read the Rundown, get per-alert verdicts with their evidence, read a record in full and add notes to it as you, check who is on call, capture operational memories, set up webhook sources with your real ingest URLs and test them through the live pipeline, and send feedback. Tools return decisions with evidence, not raw rows.

How do I connect my assistant to Signal9? Mint an Assistant, full access API key under Profile, then API Keys, point your MCP client at signal9.app/api/mcp over streamable HTTP, and send the key as a Bearer token. Narrower keys exist if you want them, and a setup-only key sees the webhook tools without the ops ones. MCP is an open standard, and Claude Code is verified end to end.

Is it safe to give an assistant access? Access is scoped by key, a read-only key cannot see the write tools at all, and anything that writes acts as you, under your permissions. Tools with real-world effects say so in their own descriptions, feedback sends only text you approve, and assistant actions can be switched off entirely.

What makes it different from other MCP servers? Most MCP servers mirror a REST API, so the assistant gets raw records and does the analysis itself. Signal9's tools decide first, the Rundown replaces alert-listing, verdicts arrive with evidence and honest abstention when history is thin, and every number carries the window it was computed over.