What it is for
Describing what you want, in plain words, and meulog configuring itself. "Tell me on Slack when a new error goes past 50 in production" is a sentence anyone on the team can say, and nobody can translate into an alert rule without opening the manual first.
What it can build
- Alert rules — who gets told, about what, with which threshold, window and silence.
- Watches — an event of your own, written as a search expression.
- Grouping rules — when to merge or split errors.
- Agents — the AI agent for a task, prompt and all.
- Dashboards — a screen with its panels: shape, metric, dimension and period.
- Settings — the values that apply to the whole meulog.
The AI writes the same file as export and import. That is not convenience, it is the boundary: the import only knows a closed list of columns per model, ignores anything else and never touches a secret. The worst a hallucinating model can do is write a silly rule — and it shows up in the diff before it is applied.
Nothing is applied without someone looking
The proposal becomes a readable diff: one line per thing, saying whether it is new or a change, and which fields it carries. The full blueprint is one click away, for whoever wants to read the YAML. Only the button applies it, and applying goes through the audit trail like any other change — with an undo.
The request is sent along with this installation's vocabulary: the channels that exist, the events, the projects, the expression fields. The model picks from what is there instead of inventing a channel that was never installed.
What it does not do
It does not install plugins, does not fill in secrets, and does not touch people or accounts. If a request cannot be met with a blueprint, it says so instead of approximating.