Settings

Settings is the single control center for Agent: it decides how the interface looks, how the chat behaves, and how the agent talks to outside services. The settings window opens from the gear icon in the app header. Each card opens its own sub-dialog on top of it — the window itself stays open behind, so you can switch between settings without closing it. Changes are saved automatically: there is no separate Save button.

Section map

The cards are grouped into five sections: General, Automation, Tools, Configuration, and Help & Account. Every card has a name and a short description; a click opens its sub-dialog.

General

Appearance — language, theme, chat font size, power-saving mode, and the simplified interface;
Chat — message display, filtering, and submit behavior;
Chat Labels — colored labels for chats: creating, editing, and deleting them. See the Chat Labels page for details.
Sound — notification sounds and volume; every sound has its own volume and a test button, including a separate auto-click sound. A separate Chat list sounds switch (on by default) decides whether the sounds of all chats are audible while you are in the chat list rather than inside an open conversation. A sound plays once, even with the app open in several tabs or windows: the other windows do not repeat it, and a badly delayed sound (a tab that was asleep) does not play at all;
Clipboard — pasting images and files into the chat.

Automation

Main Telegram Bot — set up the admin bot with all management features;
Extra Telegram Bots — set up task-specific bots with their own models and skills;
Scheduled Tasks — scheduled scripts and recurring jobs;
AutoFix — visibility of the AutoFix toggle, fixing in new chats or in the current chat, commit before the task, and choice of the AutoFix model;
AutoClick — visibility of the AutoClick toggle, autoclick during peak hours, the autoclick delay, and autocommit;
Hooks — deny and execute rules for tool calls;
MCP Servers — adding and managing MCP servers.

Tools

Analytics — usage statistics, tokens, and costs;
Proxy Traffic — inspect the chats passing through the proxy;
SSL Certificates — HTTPS certificates for the domain;
Speech Recognition — voice input with several recognition engines. See the Voice Input page for details.
Logs — server logs by day (the last 7 days, read-only): a day picker, the latest lines, and file download. See the Analytics & Observability page.

Configuration

MiniAgent — the built-in agent backend in three tabs: Profiles (named profiles), Fetch services, and Search services. See the MiniAgent page;
Retry — the interval and the number of retries for failed provider requests, stored in the retry.yml file. See the MiniAgent page;
Access Keys — named keys for programs and devices with full access to the app. See the Access Keys page;
Models — adding and managing models. See the Models page;
Container — the sizes of the mounted folders, CPU and memory load, the container’s uptime, and restart and recreate requests. See the Container and global variables section below;
Skills — managing the agent’s skills. See the Skills page;
Project Settings — editing the project’s YAML configuration files.

Help & Account

Help — open the documentation;
Support — contact support;
Logout — sign out of the application.

Simplified mode

The simplified interface is turned on from the Appearance card. In this mode the settings window shows only what you need: the Tools and Configuration groups are hidden entirely, Automation loses Extra Telegram Bots and Scheduled Tasks, and Help & Account loses Help and Support. Appearance, Chat, Sound, Clipboard, the Main Telegram Bot, AutoFix, AutoClick, Hooks, MCP Servers, and Logout stay.

Power-saving mode

Power-saving mode is a switch in the Appearance card, directly below the simplified interface. It is there for the times the app stays open for hours on a battery: several tabs with running chats drain a laptop noticeably, and the mode cuts that load. The layout does not change — not a single button moves; only how often everything refreshes.

With the mode on:

The feed updates in steps — about once every two seconds instead of continuously; what you see is always the latest version of the answer, it just arrives less often;
The answer reads as plain text while it streams — no formatting while the agent writes; headings, lists, code highlighting, and diagrams are applied once, when the answer is ready;
Tool results start collapsed — open the one you need; code highlighting runs when you open it;
Clocks and polls tick less often — the total time in the header updates in 30-second steps, quota countdowns once every 10 seconds, the chat list reconciles once a minute; the values stay correct, because they are recomputed from the current time rather than accumulated;
A hidden tab does nothing — while the tab is off-screen it refreshes nothing, and on return it catches up in one go;
Nothing moves — the tab icon does not change, the stop icon does not spin, rows appear without a fade-in, the commit confetti does not fire, and transitions and animations are switched off.

What matters is not throttled: the Stop button and the processing state, errors, the chat list and the message queue, settings, and sounds — a sound plays the moment it arrives. The clock of the current turn also ticks every second: it is how you see that the work is going.

The mode works together with the simplified interface — the two compose. Turn the switch off and everything returns to its normal behavior immediately, with no page reload.

How settings are stored

Settings are stored on the server as key-value pairs and survive a restart;
Every change is pushed instantly to all connected clients over SignalR (the live channel between the server and your browser): in another open tab or on another device, the setting updates itself;
The chat font size is remembered on the device, not on the server — each of your devices can have its own;
Power-saving mode is a device setting too, not a server one: it is kept in the browser, never sent to the server, each device has its own, and flipping it is picked up at once by the other tabs of the same device. Your other devices and the other people working in this project are unaffected — unlike the simplified interface and the language, which are stored on the server and shared;
The interface language — English or Russian — is chosen in the Appearance card: it applies at once, without a page reload, the choice is stored on the server and survives a restart, and other open windows pick it up right away. The build’s own language is used by default; the date and number formats are set by the build itself and do not depend on the chosen language;
Some keys are edited outside the settings window: the selected model and skill are set in the chat input box and become the defaults for new chats; the prompt queue settings live in the Prompts dialog;
The enforced English responses option (Hooks card) is additionally synced into the project’s local Claude settings file.

Project branding

Every project carries a small project.yml file with three fields: the project display name, the icon path, and a link to the code. The name is used in the header, the page title, and the install manifest; the icon is shown next to it. The file is created automatically on first launch, and when it changes it is re-read without a restart — the name and icon apply right away. Edit the fields in the Project Settings card.

Server updates

The client checks the server version from time to time: when it changes (after a deployment), all open tabs reload automatically — you never need to refresh the page by hand.

Container and global variables

The Container card (the Configuration group) answers questions that normally require access to the server: how much space the folders take, whether memory is running out, whether the work is hitting the CPU limit. The dialog opens on top of the settings window — nothing to ask over SSH.

What the Container card shows

Folder sizes — the data, project, home, and temporary folders: you can see at once what grows and where to tidy up. A folder missing in this deployment is honestly marked as absent, and one that could not be measured (the walk has not finished) as unavailable. The icon on a folder row opens the Usage window with the breakdown by subfolder — see the Files page;
Memory and CPU — the current usage and the limit it is measured against, plus the working set of the process itself. When the server does not run in a container, the values come from the system rather than being invented;
The container’s uptime and the moment the measurement was taken. While the dialog is open, CPU and memory refresh by themselves; close it, and the polling stops — the server stops spending time on measurements.

The dialog performs no restart, recreation, or variable save itself: it drops a request file into the exchange folder, and the managing application of the factory picks it up from the server. From inside the container there is no way to tell whether that application is there, so the dialog claims nothing about it — the sign is a single one: the file disappeared from the folder, so the request was picked up. For a plain installation on your own server the card stays useful as a meter.

Restart and recreate

Two buttons at the bottom of the dialog: Restart container and Recreate container. Both ask for confirmation first and warn that all running chats will stop; recreating additionally explains that the container will be removed and created anew from its image — the image is downloaded again, and that takes longer than a restart.

The agent restarts nothing itself: it leaves a request in the ~/global folder next to the container, and the managing application picks it up — usually within 30 seconds. After the confirmation the dialog says the request is written and the container will stop soon, stops refreshing the measurements, and tells you the page will reload by itself when the container is back. If the request could not be written, an error appears — there is no false success.

Global variables

Global variables are settings that live not in the agent but in its environment: addresses, keys, and any other values shared by every agent and factory application of this container. Handy when several products need the same address or key and you would rather not write it into each one.

The editor opens from three places: the Global variables button in the Container dialog, the item of the same name in the model dropdown of the chat input (below the model management item), and the button in the model editor — to the left of “Add variable”.

The editor is a list of name — value rows with an Add variable button and removal of each row;
A row with an empty value deletes the variable of that name; saving an empty list removes every variable set earlier. A row without a name blocks saving and is highlighted;
There are exactly two buttons: Cancel — close without writing anything, and Save and restart — write the list and close the editor;
The editor is deliberately one-way: it always opens empty and never shows the previous values. After a save the managing application picks the file up and deletes it — there would be nothing to show. For the values currently in effect, look where you set them;
After a save the variables apply to every agent and factory application of the container, and the applications are recreated — usually within 30 seconds.

Prompts — the queue of saved prompts with auto-send;
Models — peak hours and the indicator in the input box;
MiniAgent — profiles for the built-in agent;
Access Keys — keys for programs and devices instead of a single shared one;
Voice Input — speech recognition engines;
Skills — managing skills;
Hooks — security and behavior rules.

← Back: Git

Next: Skills →