# AI agents

> Let Claude or another assistant work your GiftWell for you: ask who is waiting, log a conversation, set up a new church. Each connection is named, scoped to a role, and expires.

For: Admin, Owner
Canonical: https://giftwell.app/docs/ai-agents

![AI agents: named connections, their roles and expiry, and the instructions to paste into an assistant.](/docs-img/agents.png)

## What an agent can do

Anything a person with the same role could do on the screens: list who is waiting, open a result (as bands, never a ranking), log a conversation, place someone, mark an introduction, save a ministry or a watch, find ministries for a gift, pull the gaps. For a new church it can work through setup step by step.

Everything an agent does is a note for the team. It never contacts a member, never writes to Planning Center, and in the demo church every write is refused.

## What an agent never gets

The same things a brief never gets, and more: the words of *something I have walked through*, how someone came to faith, what trusted people said in confirmations, any result or confirmation link, people who did not share with the pastoral team, anyone who may be under 18. When something exists but is withheld, the agent is told so and told to suggest the pastor ask in person.

## Connecting one

Settings → AI agents. **New connection**: a name (*Claude on my laptop*), a role no higher than your own, and how long it lives (a week by default, 90 days at most). The page shows the one block of instructions to paste into Claude Code, Claude Desktop or any assistant that speaks the Model Context Protocol. Revoke a connection any time; everything it did is in the history under its name.

## For the person running the assistant

The instructions tell the assistant how GiftWell wants results read: bands, margins, no maturity scores, no single-gift labels. If your assistant starts ranking people's gifts, that is the assistant ignoring them, and the pastor should still have the conversation.