Settings
Settings is the control center of Apis AI: which services and models answer your questions, which folders the AI may open, which extensions are available and how the app looks. Six tabs sit across the top – Providers & Models, System & Prompts, Extensions, Parameters, Appearance and Legal & Help. Nothing takes effect until you click Save Settings; a short Changes saved note confirms it.
Providers & Models
This is where you set up the services that actually answer – your own Ollama on the computer, or a cloud service such as OpenRouter, Anthropic, DeepSeek or Puter. Each service appears as its own card.
Provider & API
On a provider card you fill in how Apis AI reaches the service and how it should talk to it:
| Field | What it is for |
|---|---|
| Provider | The type of service – OpenAI-compatible or Anthropic (Claude). Predefined providers set this automatically, because the type decides the request format. |
| Name | How the provider should be called in chats and in the sidebar. Free choice. |
| Base URL | The web address the requests go to. For a local Ollama this is your own machine. |
| Context Window (Tokens) | How much text the model can keep in view at once. With Recognize Apis AI fills the value in automatically; you can also enter it yourself. |
| API Key | The access key of the service. A missing key is flagged directly on the card, since some providers cannot load any models without it. |
| Thinking | Lets a model "think" before answering. Meant for models that support this mode. |
| Max Tokens Field | How the maximum answer length is passed on. Automatic detects the suitable field on its own – that is the right choice for almost every provider. |
Advanced Settings hide the finer details of the connection. With Add a provider you create a new card, with Remove you delete one. While a provider is loading its models you see Loading models…; if the list stays empty, a missing or wrong API key is the most common reason.
Choosing models
The model dropdown on a provider card lists everything the service offers; in a long list you can search by name (Search for a model…) or refresh the list. The star marks models you use often: Add to Favorites and Remove from Favorites keep them at the top. Models that cannot call tools are marked No Tool Call Support – they can chat, but they cannot use extensions, skills or file access.
Each provider's model list is loaded once and then cached, so reopening the dropdown does not
query the service again. The list is refreshed automatically as soon as you
change the provider's endpoint, API key, kind, GGUF folders or auto-start. In the chat's
"+" menu the refresh button in the model list header forces a new query at any
time – for llama that re-scans the configured folders, so newly added .gguf files
appear without restarting the app.
Standard Provider & Model
Here you decide what happens when you start a new chat: the Default Provider and the Standard Model are used automatically. The sidebar and the model field in chat show your current default at a glance.
System & Prompts
Three things live here: the version of Apis AI you are running, which folders the AI may look at, and what you want to tell it about how you like to work.
Updates & Version
The card at the top of the tab tells you which version of Apis AI is installed and whether a newer one exists.
| Element | What it does |
|---|---|
| Installed Version | Your current version number, for example v0.1.4. Open the website takes you to the download page in your browser. |
| Check for new versions automatically at startup | While this switch is on, Apis AI asks once at every start whether a newer version exists. Turn it off if the app should not go online on its own. |
| Check for updates now | Checks immediately, without restarting the app. The result appears right next to the button. |
After a check, the line next to the button tells you where you stand:
| Message | Meaning |
|---|---|
| Check for updates… | The app is still asking. A check takes a moment. |
| Apis AI is up to date (v…) | You are running the latest version – nothing to do. |
| New version v… is available! | A newer version exists. You download it with Open the website. |
| Error checking for updates: … | The check did not get through, usually because there is no internet connection. Your app keeps working normally – simply try again later. |
When a newer version exists, Apis AI also points it out by itself: as a hint above your chats with Open the website and Close, and as a small Update v… badge in the sidebar above the chat list. Both are only reminders – the app never replaces itself. How a new version gets onto your computer is described under Installation.
Shared Folders
Apis AI only reaches folders that you release here – and nothing else. With Add Folder (Home) a folder picker opens in your home directory; with Remove you take an entry out of the list again. Without a released folder the AI cannot read or write any files. What the AI is allowed to do inside these folders is described under Tools & Security.
System Prompt
Under Custom Instructions (optional) you add your own wishes, for example that answers should always be short and in your language. Apis AI keeps a technical base prompt of its own, which is always sent first and cannot be removed. If you want to see what the model really receives, Show full prompt unfolds the whole text.
Speech Input
The Speech Input (Speech-to-Text) card enables voice input. It is off by default;
once enabled, a microphone button appears in the chat input. Choose between a local
Whisper.cpp server and a remote OpenAI-compatible service, then click
Test connection. For the local server you can switch on
Start Whisper server automatically, so Apis AI launches whisper-server
itself at every app start instead of you starting it by hand. The button
Guide in the wiki opens the detailed setup page. Full instructions:
Speech Input.
Rendering
Hardware rendering (experimental) is only relevant on Linux with Wayland. Off
(the default) Apis AI uses software rendering there – the reliable choice with the
proprietary NVIDIA driver, where the GPU path sometimes shows a black or frozen window.
If you switch it on, the app tries the GPU renderer; if the GPU delivers no image within
a few seconds, Apis AI automatically falls back to software rendering and tells you about
it in the chat. An environment variable SKIKO_RENDER_API always takes
precedence over this setting.
Extensions
Extensions give the AI additional capabilities. This tab contains four cards:
| Card | What it is for |
|---|---|
| MCP Server | Connections to further services – calendars, mail, project tools and much more. How to add and use them: Understanding & Using MCPs and for the technical setup Setting Up MCP Servers. |
| Tool Selection per Model | Decide per model which MCP servers and which internal tool groups are sent with a request – the remedy for coding models with a small context window. Details: Tool Selection per Model. |
| Rules | Permanent instructions that apply to every chat. |
| Skills | Reusable capabilities that only take effect when they are needed. |
Rules and skills are explained step by step under Rules & Skills.
Tool Selection per Model
Every tool the AI may use is announced to the model with every request – with its name, its description and its parameter schema. That list is sent before your message and takes up part of the context window. The more MCP servers are connected, the longer it gets, because a single server often brings a dozen tools along. Coding models with a small context window are overwhelmed by that: they answer without using a tool at all or the request fails outright – and then neither reading nor editing files works.
The card Tool Selection per Model keeps a list with one entry per model, several models can be configured next to each other. Add model below the list appends a new entry – you pick the provider and the model right inside it, and as soon as the model is chosen the entry unfolds with the switches for what this model should not receive:
| Switch | What it does |
|---|---|
| Send MCP tools | Master switch for all MCP servers. Off = this model receives no MCP tools at all. On = every server gets its own switch. Servers you switch off here are hidden from this model only; servers switched off globally on the MCP Server card stay off for everyone. |
| Internal tool calls | Files & folders, terminal, scripts, task list, charts, math, skills, questions, chat title, history archive and screen. Each group lists its tool names in the description, so you can see exactly what disappears. |
A switched-off group disappears twice: its tools are no longer sent, and its instructions are removed from the system prompt, so the model does not even try. If it still calls such a tool (for example because it saw it in an earlier message), the call is not executed – the model gets a short note that the tool is disabled for it. Since Files & folders is what makes file editing possible in the first place, the card warns you when you switch it off.
Every entry you save shows up in the list as an expandable card with a one-line summary, for example Tools: 2 of 5 MCP servers · without Charts, Screen. A click on the header unfolds the switches again, Remove entry deletes the entry. An entry is stored per provider and model name – capitalisation does not matter – and lives in your settings file next to providers and shared folders. Models without an entry keep receiving all tools; an entry without any restriction says exactly that and stays in the list, so you can keep the models you care about together and extend them later. Picking a model that already has an entry opens that entry instead of overwriting it.
In a workflow, the model profile and the selection on the node (Agent Node) are combined: a tool has to be allowed in both places to be available.
Parameters
Everything that concerns the working style of the AI, the app around it and the size of your chats is collected here.
Security & File Operations
| Option | What it does |
|---|---|
| Can delete files on its own | On: the AI deletes files without asking. Off: you confirm every deletion beforehand. |
| Terminal Sandbox (shell_cmd) | On (default): the AI may only run allowed, read-only commands. Off: full system access, any command. |
| Whitelist / Blacklist | Applies while the sandbox is active. Whichever mode you choose decides whether only the listed commands are allowed or all commands except the listed ones. One command per line; an empty list means the default of that mode. |
| Command timeout (shell_cmd) | How long a single command may run, in seconds (1-1800, default 30). Once it is reached, the command and all of its child processes are stopped and the output read so far is reported. The AI may ask for a shorter limit per call, never for a longer one. Work that legitimately takes minutes belongs in a script (600 s). |
| Reset to System Default | Discards your own command list and restores the delivered one. |
A fixed safety baseline stays in place in the sandbox: destructive commands and interpreters are never run, not even if they appear in a list. To allow them you would have to switch the sandbox off completely. For individual chats and workflow nodes you can decide differently. More on this under Tools & Security.
Tasks & Reminders
| Option | What it does |
|---|---|
| Reminder of Unfinished Tasks | On: while tasks are still open, Apis AI reminds the AI to keep going. Off: the AI finishes when it thinks it is done. |
| Ask about tool calls | The fallback interval. It decides how many tool calls may pass since the last check before the AI is gently asked whether everything is finished. 0 switches the fallback off. The count starts at the last check, not at the start of the chat. |
| Tool Failure Reminder | On: if the same tool keeps failing, Apis AI tells the AI that every attempt fails and that it should change its method instead of repeating the call. This prevents endless tool-call loops. |
| Remind after failed attempts | How many failures of the same tool in a row may occur before the reminder is sent. 0 switches the reminder off. Values below 2 are treated as 2, so a single failing attempt never triggers it. A successful call starts the count over. |
Chat Input
Press Enter to send the message decides how you write: switched on, Enter sends and Shift+Enter inserts a line break. Switched off, Enter inserts a line break and Ctrl+Enter sends.
Time & System Info
With Send time & system info every request carries the current date, the time and some information about your computer (operating system, Java, locale, user) as metadata. That is what makes questions like "What is on today?" or time-based workflows reliable. Off (the default) means: no environment information leaves your computer.
Background Service
| Option | What it does |
|---|---|
| Run in the background (system tray) | Like Discord, Apis AI keeps running in the tray when you close the window. Scheduled workflows then run even without a visible window. |
| Start automatically when logging into the system | Apis AI starts together with your computer, so scheduled workflows also survive a reboot. |
| Notifications for Scheduled Workflows | Off, Only if there are errors or Success and Mistakes – you decide how much you want to hear about automatic runs. |
Details on the tray, automatic starts and schedules are in Background Service & Schedules.
Workflow Context
Max. Handover Characters (HANDOVER) limits how much text a handover package may contain when one workflow step passes its result to the next. Packages that are too large are shortened. A value set in a single node or in the workflow itself wins over this default.
Model Parameters
| Option | What it does |
|---|---|
| Temperature | How creative the answers may be. Low values stay sober and factual, higher values are more varied. |
| Max. Tokens | The upper limit for the length of a single answer. |
| Max. Tool Rounds | How many steps the AI may take in a row before it has to stop. 0 means no limit – it works until it is done or you stop it. |
| Compression Threshold | The share of the usable input budget at which a long chat is automatically condensed. The usable budget is the context window minus a reserve for the answer (min. 20,000 or Max. Tokens, whichever is smaller), so the answer always keeps room. While compressing you see Compressing chat… at the end of the conversation (for every provider; shown only if the compression takes more than a moment). |
| Messages After Compression | The upper limit of messages that stay verbatim after a compression. Older ones remain visible, but are no longer sent to the AI. |
| Tokens After Compression | The token budget for the verbatim tail. 0 = automatic (min. 15,000, max. 2,000, 25 % of the usable budget). At least two messages always stay. The message count remains the upper limit. |
| Prune Old Tool Outputs Before Summarizing | Before a summary, old tool results (outside the last ~40,000 tokens) are replaced by their handle if this frees more than 20,000 tokens. The full output stays readable via tool_result_read. Off by default. |
Compression and the context meter are explained in Getting Started.
When is compression triggered?
The trigger is based on the usable input budget – the context window minus the reserve for the answer. The app also counts the real token usage reported by the provider, so a chat is compressed in time even when the internal estimate under-counts. If the provider still rejects a request because the context is too long, the app drops media from older messages and compresses once, then repeats the request exactly once.
What stays in full length:
- the first user message and the pinned active task (anchors) – they are never folded and are re-injected after every compression;
- the most recent messages up to the token/message budget;
- all previous handover reports – they are never summarized again, so no knowledge is lost through repeated summarising.
The handover report has five fixed sections. A report without any usable content (no section with substance) is rejected, the history stays unchanged and the reason is shown in the chat. Individual sections that legitimately have nothing to say are allowed. The full original history remains available to the AI through the archive tools (history_search, history_read) and to you through Rewind.
Appearance
Here you make Apis AI look the way you like – this only affects the program, never the answers of the AI.
| Option | What it does |
|---|---|
| Theme | The colour world of the app: Seafoam, Ocean, Sunset, Forest, Berry, Monochrome, Lavender, Midnight, Sakura and Gold. |
| Mode | Dark for a dark interface, Bright for a light one – whichever is easier on your eyes. |
| Language | The language of the interface. |
| Global scaling factor | Font size of the whole app, from 80% to 150%. The change is visible immediately after saving – handy on high-resolution screens. |
| Chat Title Mode | How a new chat gets its name: AI Generates Titles has the AI suggest one, First User Message simply takes the beginning of your first message. |
Legal & Help
The last tab has nothing to configure – it is there for you when you need something in writing or want to get in touch.
- Legal Information & Licenses – with Open Legal Information the legal notice, the privacy policy and the licences of third-party components open in a window, even without an internet connection.
- Support & Help – with Contact Support you write an e-mail to our support team, directly from the app.