AI agent
Agent chat basics
Right panel → ask a question → agent calls tools and replies.
What the agent is
The Kition Agent is not a chat box — it is a local AI workmate that can read your vault, edit files, run commands, and drive a browser. Orchestration happens on your machine; only the model call leaves to a provider.
It can read your vault (docs and tables) out of the box. Writes, Bash, and network fetches are gated per-tool by the permission model — you stay in control.
What one chat can do
- Read any doc / table / attachment in the vault
- Fetch web pages, summarize, cross-compare sources
- Run Bash (gated by both hooks and permissions)
- Modify docs, add table rows, write new files
- Drive a real browser through logins, screenshots, form fills
- Mount MCP servers for GitHub / Jira / Slack and more
Open a chat
Cmd/Ctrl + Jtoggles the panelCmd/Ctrl + EntersubmitsShift + Enterinserts a newline@references a doc / table / view as context/triggers slash commands (plan, goal, subagent, ...)
A typical opener
The sharper the goal, the better the tool selection. "Help me with docs" leaves it guessing; "Read @projects/onboarding.md, pull the TODOs into a table with owner and due columns" lands in one shot.
The agent is bad at guessing implicit constraints. Stating "concise tone, under 200 words, no emoji" upfront is far cheaper than interrupting and redoing mid-flight.
Sessions and history
Each chat is its own session, stored under .kition/sessions/ in the vault. Fork an old session, tweak the prompt, and rerun to compare strategies side by side.
When context exceeds the model window, the agent compacts older turns into a summary and keeps recent detail. Tune the threshold in Settings → Agent → Context.