Use Averil from your AI tools
Averil's MCP server lets you ask about your security cases from Claude Code, Cursor and VS Code, without opening Slack.
What it does
All three speak the Model Context Protocol (MCP). Connect them to Averil and ask things like "Anything I should know before I touch the billing service?" Your AI tool calls Averil's read-only tools and answers from your cases.
The tools
| Tool | Returns |
|---|---|
security_brief | What opened, what was verified fixed, what was dismissed and what waits on people, over 1 to 14 days, with your own cases first |
list_cases | Open cases, the ones that need a person, closed or all, optionally only yours, up to 50 |
get_case | One case with its latest investigation summary, evidence references, proposed plans and a console link |
case_timeline | Everything that happened on a case, in order, up to 100 entries |
resolution_metrics | Verified-resolution rate, owner-confirmed share, reopen rate and time to verified resolution, over up to 90 days |
Every tool is marked read-only. None of them can approve a plan, move a case or change one of your systems. Approvals stay in Slack and the console.
Connect a client
- 1In the console, open AI tools and create a personal token. It is shown once and lasts 30 days.
- 2Run the command the console shows, or in Cursor and VS Code add a remote MCP server with the address and an Authorization header.
- 3Ask a question. The client calls the tools for you.
In Claude Code the command looks like this:
claude mcp add --transport http averil https://mcp.averil.ai/mcp --header "Authorization: Bearer <your token>"Sign-in and limits
- Your token belongs to you and works only with the MCP server, not the rest of Averil.
- Only people can call it. Averil's own agents and service credentials are refused.
- Your company comes from your token, never from what the client or its model sends.
- Each person can make 60 requests a minute.
- Every call is recorded in your action log with the person, the tool and the case.
What it never returns
Text is redacted before it leaves, with the same rules that guard model prompts. Evidence is returned only as references (ID, source, SHA-256 and time), never its content. Case text can quote logs and code, so every result carries a notice that quoted text is data, not instructions.
Coming soon: sending a diff for Averil to review and prove, and drafting a fix plan that then goes to the case thread for approval.
Stuck? Email averil@averil.ai or ask me in Slack.