Skip to main content

I want to enable Nightingale AI

Configure an LLM provider, turn on the built-in assistant and the MCP endpoint, and connect a client such as Claude Code.

Where this path ends: you can ask "which S1 alerts are firing right now" inside Nightingale in plain language, and an MCP client such as Claude Code can ask the same thing. About 10 minutes.

Before you start​

  • Nightingale is running;
  • an API key for a model provider, or an Ollama running locally.

1. Configure a model​

The entry point lives inside Nightingale AI (the top-level item in the sidebar); its own secondary navigation has LLM configs. There is no top-level menu for it — visiting /ai-config/llm-configs redirects here.

Click Add LLM config and fill in:

FieldNotes
NameYour own
Provider typeThe form offers three: OpenAI compatible, Anthropic Claude, Google Gemini
API URLOpenAI-compatible services (vLLM, Azure OpenAI, Ollama, Kimi…) all use OpenAI compatible with your own URL
API key
ModelThe model name, e.g. gpt-4o, claude-sonnet-4-5
Context length

Save it and mark it default: anything that has no model of its own falls back to this one.

2. Ask it something​

Back in Nightingale AI, start a chat and ask:

Which S1 alerts are firing right now?

The assistant queries your event data before answering. A useful answer means both halves — model and data — are wired up.

3. The MCP endpoint is already on​

/mcp and /a2a are enabled by default; no config change needed. They reuse HTTP.TokenAuth, so a client authenticates with X-User-Token. They also accept Authorization: Bearer when the built-in authorization server or external-IdP auth is enabled.

The important part: an MCP client has exactly the permissions of the token's owner, and no more. Create a dedicated least-privilege account for AI access rather than handing over root's token.

Only read tools are registered by default. Write tools (create / update / delete) are an explicit opt-in:

[HTTP.A2A]
MCPEnableWriteTools = true

You can also expose a subset of toolsets:

# empty = all default toolsets
MCPToolsets = ["alerts", "dashboards"]

There are 13 toolsets: alerts, targets, datasource, mutes, busi_groups, notify_rules, alert_subscribes, event_pipelines, users, metrics, logs, dashboards, roles. Unknown names are ignored — the list never widens by accident.

4. Connect a client​

Create a token first: avatar menu → personal settings → generate a token. Then point the client at Nightingale's /mcp endpoint with that token.

Verify with something read-only, such as "list all alert rules".

Next​

On this page