Settings
The Settings tab and its nine pages — and the thing worth knowing first, which is that they don't all remember things in the same place.
The Settings button at the right end of the menu bar, or Ctrl+,.
It opens as a tab in the editor strip, like a file, and takes the whole centre of the window.
Both side panels slide away while it's in front; switch back to any other tab and they return.
Ctrl+, opens it, or jumps to it if it's already open — it never closes it. Close it the way you
close a file tab.
Down its left edge is a rail of pages in two groups:
App — Studio, Graph & Explorer, Keybindings, Privacy, Project.
Agent — Models, Tools, MCP Servers, Instructions.
Settings reopens on whichever page you used last. Pages stay loaded once you've visited them, so a long-running action on one page (a re-index, say) keeps going while you read another.
Every switch, stepper and dropdown applies the moment you change it — there is no Save button for
those. Things you create — a model, a server, an instruction, a tool — are edited in forms with
Save (or Create) and Cancel, and nothing you type there is stored until you press it. Nothing
anywhere in Settings needs a restart.
Note
Settings needs an open project. On the startup screen there is no menu bar and no tab strip, so the shortcut does nothing — open a project first.
Where your settings live — the part that surprises people#
The nine pages look like one thing. They are not: what you set here is stored in five different places, with five different ideas of who it belongs to. Which one a setting uses decides whether it follows you to your next project, to your next machine, or nowhere at all.
Per project. Everything on Graph & Explorer, all of your Keybindings, and Privacy's
Walk memory every number. These are remembered beside that project's index, so they're remembered
per project and only per project. Rebind Ctrl+J here and it's still Ctrl+J in the next project
you open.
Per machine, every project. The Studio page — the emoji picker and Semantix's own model
seats — plus your connected providers, your custom models, your MCP servers, your
global instruction, your user-scope slash commands, your native tool switches and your
user-scope custom tools. These live under ~/.semantix/USER/, so they apply in every project you
open on this computer — and they don't travel to another one.
Per project, as files you can edit. Project instructions, workspace slash commands and
workspace custom tools — markdown and JSON under ~/.semantix/WORKSPACE/<project-id>/, readable
and editable outside the app.
Per account, everywhere. Only the three switches on Privacy. Those are stored against your account, so they follow you to any machine you sign in on.
Somewhere else entirely. Fonts, sizes, the editor's syntax theme and every colour are on the Theme tab, not in Settings at all — see below.
The Instructions, MCP and Tools pages tell you their scope on screen — a Workspace / User toggle,
a ws / usr badge, the storage path printed under the heading. The others don't say anything,
which is why this is worth reading once. When a setting mysteriously "didn't stick", it almost
always stuck — in a scope you weren't looking at.
The nine pages#
Studio — the composer's emoji picker, and the two model seats Semantix uses for its own background work: building your memory, carrying context between sessions.
Graph & Explorer — the switch that turns the 3D view off entirely, and whether clicking a folder in the explorer unfolds it inside the main graph or opens it as a graph of its own.
Keybindings — see the app's shortcuts and rebind about half of them; the rest are listed but fixed. Also worth reading for the shortcuts that work but aren't in there at all.
Privacy — what gets stored, what gets backed up, whether the agent remembers you between conversations, and what you can turn off.
Project — the Re-index button, which is the fix for the IDE having missed changes you made while it was closed.
Models — connect a provider with one key and get its whole catalogue, switch the coding agents installed on this machine on and off, or point Semantix at any OpenAI-compatible or Anthropic-compatible endpoint of your own.
Tools — which of the built-in tools the agent gets, and your own custom tools: a bash script or an HTTP call with typed parameters, called by name like any other.
MCP servers — connect external tool servers. Anything they expose becomes
available to the model on every lane, named mcp__<server>__<tool>.
Instructions — standing instructions that ride along with your
messages, and reusable /slash prompt templates.
What isn't here#
Theme, colours, fonts and sizes. The Theme button sits just left of Settings in the menu
bar and opens its own tab on the same rail-and-page shell, with seven pages: Presets, Colours,
Code, Typography, Glass, Wallpaper and Section CSS. The interface font and size, the chat
size, the editor's typeface, size and syntax theme, and the minimap switch all live there.
Choosing a model for a conversation. The Models page is where you make a model available; the picker in the chat composer is where you choose one, per conversation.
Your Claude sign-in. Nothing in Settings touches it. The Claude Code card on the Models page
only tells you whether the claude command on this machine is signed in.
Permissions. How much the agent may do without asking is decided by the Manual / Edits /
Full chip next to Send, not by anything in Settings.