AI Features

How AI Works Here

Every AI feature is optional and switched off until someone adds a connection. There is no Kwilio model. You supply a key for a hosted provider and that provider bills you directly, or you point the app at a model running on your own machine and nobody is billed at all.

Two rules shape all of it:

• Nothing runs on its own. Generating costs money, and a page load is not consent. You ask, and it answers.
• Every answer is reconciled against your actual board before you see it. What doesn't check out is reported back to you rather than quietly patched.

Adding a Connection

Add an AI connection from the organisation settings. You can use a hosted provider with your own API key, or a local model.

Your AI keys are read in exactly one place on the server and are never sent to the desktop app — the app can replace a key, but never read one back.

Running a Model Locally

Point Kwilio Projects at LM Studio or Ollama on your own machine. The text never leaves your computer, nothing is billed, and a spend cap of zero still leaves local models working.

What It Can Do

Draft a specification

From a one-line topic to background, requirements, acceptance criteria and test scenarios. The result is stored as structured data, so it can become subtasks without asking the model again.

Triage before you file

A priority, labels and an estimate for a task that does not exist yet. Invented labels are dropped, estimates over 80 hours are refused, and duplicates are found with trigram search rather than a guess.

Release notes that reconcile

"Done" and "merged" disagree more often than you'd like. Work ships if either is true, and every disagreement comes back to you as a finding rather than being quietly resolved.

Search in plain English

"Overdue design work in the web project" becomes a query, and Postgres answers it. The model never sees a task and never names one, so a hallucinated result is impossible.

Does the code match the spec?

Reads the branches and pull requests attached to a task against what the task promised. "Unclear" is a first-class answer — a confident wrong verdict about finished work is worse than an honest shrug.

The dashboard brief

Your open work across every project, ranked, with a reason for each — because a sorted list can't tell you that a two-minute review is holding up a colleague.

• Code picks the candidates and states the facts; the model only orders them
• Overdue work the model left out is reported back to you, not quietly dropped
• Capped at five, because a longer list is the backlog you were escaping

Privacy

Prompts go to the provider whose key you supplied, or to your own machine, and nowhere else. There is no Kwilio model to train, and your work is never used to train one.