# 88 — LLM agent chat

## Goal

Natural-language chat that acts on the app via tool-calling against the API, using a provider-agnostic OpenAI-compatible endpoint (self-hosted on the homelab).

## Why

Users should say "add milk to the shopping list" or "what's on the calendar Friday?" and have the agent use the API.

## Depends on

- 87 (API surface)
- ADR `0006-api-llm-agent.md` agent half (MCP yes/no, tool mapping)

## Scope

- Agent chat UI in the app.
- OpenAI-compatible client (provider-agnostic).
- Tool-calling mapped to API endpoints from 87 (and journal if present).
- Decide MCP vs direct tools in ADR 0006.
- Smoke tests with **mock/stub** provider in CI — no live LLM in CI.

## Out of scope

- Training or fine-tuning models.
- Forward-auth for other homelab apps.

## Acceptance criteria

- [ ] Chat UI sends prompts and shows responses.
- [ ] Tools call the API successfully in dev against the homelab endpoint.
- [ ] CI uses mock provider only.
- [ ] ADR 0006 records provider and MCP decisions.

## Notes

- Gitea: [#16](https://gitea.ginnoir.com/ginnoir/famapp/issues/16), [#17](https://gitea.ginnoir.com/ginnoir/famapp/issues/17), [#18](https://gitea.ginnoir.com/ginnoir/famapp/issues/18)
- Depends on [#15](https://gitea.ginnoir.com/ginnoir/famapp/issues/15)