Exe AI — Manual

Chapter 01 / 16

Getting started

From the first launch to your first response.

When you first launch the app, you see the Welcome window. It asks for three things: your preferred name, language, and appearance. You can change all of these later in Settings. You do not need an account.

Three steps to your first model

Step 1

Download the server software

Open Local at the bottom of the left sidebar. Use the download button to download the model server: the small, free program that does the computing.

Step 2

Choose a model

Open the Catalog. It lists models that fit your computer's memory. Select Download to download your chosen model.

Step 3

Start

Go back to Local and press Start. Context, GPU layers, and port are already filled in. When the indicator turns green, the server is running and your model is available in the chat.

Prefer a cloud model? Open Cloud, select Add provider, enter your key, and press Connect. The provider's models appear automatically.

Chapter 02 / 16

The window

The sidebar is on the left, the chat is in the middle, and a second sidebar on the right holds everything created alongside the chat. Both sidebars are collapsed after installation.

The Exe AI window with chat and sidebars — Inside Exe AIEnlarge image ↗

⌘ B · Ctrl + B

The left sidebar

At the top, you will find Search chats, New chat and New project. Below are your projects and chats, sorted by date. At the bottom, you will find six icons:

Chats	your conversations
Local	the servers on your computer
Cloud	providers you connect to using a key
Catalog	find and download models
Settings	everything else, including this manual
Images	generate images and videos

⌘ J · Ctrl + J

The right sidebar

This shows what the assistant creates and does alongside the chat. There are six icons at the bottom here, too:

Preview	files the assistant creates: HTML, SVG, spreadsheets, PDFs
Agents	start tasks, track progress, and read reports
Cron jobs	a calendar for everything scheduled to run at a set time
Terminal	every command in this chat, logged live
Gallery	all generated images
Doc folder	uploaded and saved files

In the preview, use View and Source to switch between the finished result and the file's text. Reload refreshes the file; Download saves it.

The status bar shows whether the model server is running — Server active or Server inactive — and how much memory is in use.

Chapter 03 / 16

Chats and projects

Each conversation is a chat. Projects keep related work together.

New chat starts an empty conversation. Search chats helps you find an existing one. Right-click a chat, or click the More button in its row, to open its menu:

Pin	Keeps the chat at the top, under “Pinned.”
Rename	Gives the chat a new title.
Move to project	Moves the chat into a project.
Archive	Removes the chat from the list without deleting it. In the Archive, use Restore to bring it back.
Delete	Deletes the chat. If an agent is still working in it, the agent is stopped as well.

Projects

A project is a named group of chats. Chats in the same project share their working folders. New project creates a project. You can also archive an entire project.

The indicator next to a chat

Yellow	Waiting for you
Blue	Response in progress
Green	Response complete
Red	Run failed

Chapter 04 / 16

The input bar

Below the text field is a row of buttons. They apply to the chat that is currently open.

Upload files	Attaches documents and photos. You can also drag or paste files into the field.
Working folders	Grants access to folders where the assistant may work.
Permission mode	Auto, Manual, or Restricted: controls how often the assistant asks before using a tool.
Model	Selects the model for this chat. Three indicators show whether it can call tools and see images, and whether its speed module is active.
Context ring	Shows how full the context is, for example “0 / 16k tokens.” Context is everything the model can consider at once: the chat so far, your documents, and the tools.
Thinking	Sets how thoroughly the model thinks before responding: Off, On, Minimal, Low, Medium, High, X-High, or Maximum. The available levels depend on the model. Some show only “Always thinks” or “Never thinks.”
Dictate	Click to start dictating into the text field. Hold the button for one second to start a conversation.
Send	Sends the message. If a response is being generated, the same button stops it.

Keys in the text field

`Enter`	sends the message
`Shift + Enter`	starts a new line
`Escape`	stops the response
`/`	as the first character: shows your skills

The queue

If you write while a response is being generated, your message is marked Queued. It is sent as soon as the response is complete. Remove from queue returns it to the input field.

Chapter 05 / 16

Reading responses

A response shows both the text and how it was produced.

Thinking

If the model uses reasoning, a collapsed row labeled Thinking appears above the response. Click to expand it.

Tools

Each tool call appears as a small card showing its name and duration. If the call fails, the card is marked “Failed.” Click to see exactly what ran.

Below the response

Copy copies the text to the clipboard. Regenerate asks the model to respond again. You will also see performance metrics: TPS shows how quickly the model generates text; TTFT shows how long it took for the first text to appear.

Follow-up questions

Sometimes the assistant asks you a question. Choose one of the answers or write your own. Answer later postpones the question.

When the context fills up

When the context is almost full, the chat summarizes its older messages. A label such as “12 compressed messages” appears in their place. Nothing is lost: the history remains readable, while the model receives the summary. Set the threshold under Settings → Instructions → Context compression. “Auto” means 80%. This feature is still in development.

Chapter 06 / 16

Models

Local means that nothing you type leaves your computer. Cloud means that your text is sent to a provider, which bills you for usage.

The catalog

The catalog has six tabs: Chat, Embedding, Speech, Image, Accessories, and Downloads. Each card shows a model's capabilities: “Calls tools,” “Sees images,” “Thinks before responding,” and “Has a speed module.” A line such as “Recommended for your 64 GB” tells you whether it fits your computer. You can also use the search field to find models directly on Hugging Face. Under Downloads, you can see what is currently downloading.

Local: six servers

Under Local, each server has its own card.

Model server	The model you chat with.
Agent server	The model used by agents and cron jobs.
Embedding server	Prepares documents for search. It is loaded on demand.
Speech server	Dictation, conversation, and voice.
Cleanup server	A small model that polishes dictated text.
Image server	Keeps the image model loaded so that image generation does not have to start from scratch each time.

On a server's page, choose a model and configure it: Context (how much the model can consider at once), GPU layers (how much of the processing the graphics card handles), CPU threads, plus the Speed module and Vision module. Under Advanced, you will find settings such as KV cache, Flash Attention, and Memory Lock. The Server log shows the server's messages. Use Start and Stop to start and stop it.

Under Parameters, you will find values such as temperature. They apply to all chats using this model. Reset restores the defaults.

Cloud

Add provider connects Anthropic, OpenAI, or a custom server that supports the OpenAI API, such as oMLX or LM Studio. Your key stays on this computer, in a file read only by the app. After that, the interface only shows whether a key has been set.

Cloud providers often do not report a model's capabilities. Enable the appropriate options yourself on the model's page: Model can see images (Vision), Model can think (Reasoning), Model can use tools (Tool Calls).

Chapter 07 / 16

Folders and documents

In an approved folder, the model works without asking. For access outside that folder, it asks in the chat.

Working folders

Use Working folders to grant access to a folder: browse for a folder or paste its path, then select Grant access. Grant access to another folder adds another one. Revoke folder access removes permission without deleting anything.

Each chat also has its own folder. Chats in a project share the project's folders.

Folder notes

Each approved folder contains a file named exe.md with notes about that folder. These notes accompany every message, and the assistant keeps them up to date.

Attaching documents

Use Upload files to attach images, for models that can see them, and documents up to 20 MB: PDFs and text files such as TXT, MD, CSV, JSON, or source code files.

A document is included in full in the context; its card then says “Fully in context.” If it does not fit, the card says “Document too large.” The file still remains in the chat's folder, and the assistant reads individual passages from it.

The Doc folder

The Doc folder in the right sidebar collects everything you have uploaded and everything agents and cron jobs have saved. Each file shows who created or uploaded it. The All projects view shows files from every project.

Drag a file into the chat to attach it. Add files adds new files; Change location moves the folder.

Chapter 08 / 16

Tools and permissions

Tools let the assistant do more than write: read files, run commands, search the web, and generate images. You decide what it is allowed to do.

The MCP server gallery — Inside Exe AIEnlarge image ↗

Permission mode

Auto

Tools run without asking, even outside working folders.

Manual

File and command tools work without asking inside working folders. Outside those folders, and for other tools, permission is generally requested first.

Restricted

Permission is requested before every file operation and command, even inside working folders.

After a fresh installation, the mode is set to Manual. Until you choose a mode yourself, the app asks only about file operations and commands outside working folders. For agents, the middle level is called “Sandboxed.”

The “Allow execution?” prompt

When a tool needs your permission, the prompt shows exactly what will run and why approval is needed, for example because a path is outside the approved folders or the operation accesses the network.

Allow once	Applies only to this particular call.
Always allow	From now on, applies to the entire tool in all chats and all agent runs.
No	Denies permission. You can add a note explaining what the assistant should do instead.

To revoke an “Always allow” permission, edit berechtigungen.json in the app's data folder (see Your data). Some dangerous commands remain blocked in every mode.

Unattended agents reject anything that would need approval. Agents you start from a chat send their approval requests to that chat.

Built-in tools

Under Settings → Tools, each tool has a switch. Every enabled tool adds a little time to each request, so disable the tools you do not need.

Group	Tool	What it does
Terminal	`read_file`	Reads the contents of a text file.
	`write_file`	Writes text to a file, creating it if necessary.
	`edit_file`	Replaces an exact match in a text file.
	`list_dir`	Lists the contents of a folder.
	`search_files`	Searches approved folders for text and reports every line containing it.
	`create_folder`	Creates a folder and grants access to it as a working folder for this chat.
	`run_command`	Runs shell commands on this computer, for example to start programs and scripts or move and rename files.
	`ask_user`	Asks you when a decision needs your input.
	`start_job`	Delegates work to an agent that completes it in the background and leaves a report.
Exe Web Search	`web_search`	Searches the web.
	`read_webpage`	Reads a web page.
Memory	`memory_save`	Remembers information about you or your project across conversations.
Skills	`skill_load`	Loads saved instructions for a type of task.
	`skill_save`	Records a workflow so it can be loaded again later.
Image tool	`draw_image`	Generates an image using the local image model.
	`draw_video`	Generates a video using the local video model.
Cron jobs	`cronjob`	Sets up one-time or recurring reminders that trigger a task at a set time.
	`cron_note`	Saves a short note for a scheduled event to be read again during that event's next run.
Clock	`get_current_time`	Reads the current time, day of the week, and time zone.
Goals	`goal`	Notifies you that an agent has reached its goal.
Dock	`deliver_file`	Saves a finished file in the Doc folder, where you can open or download it.

More tools through MCP servers

Under Settings → MCP servers, connect Exe AI to apps you use every day, such as Notion, GitHub, Figma, Gmail, Recraft, Hugging Face, Google Drive, or Slack. Depending on the service, sign in through your browser or enter an API key. Under Custom server, enter an address or launch command. Web search is already built in.

The app treats tool output as data, not as instructions. This also applies to text from web pages.

Chapter 09 / 16

Agents and cron jobs

An agent completes a task independently and returns a report. A cron job runs a task at a scheduled time.

The agents panel with tasks and reports — Inside Exe AIEnlarge image ↗

Creating an agent

Settings → Agents. An agent has a name, description, instructions, tools, folders, and limits: Max. rounds, Max. minutes, Max. tokens. Each agent is stored as a single Markdown file.

Starting a task

Right sidebar → Agents → New task: choose an agent, describe the task, and select Start. You can also ask the assistant in the chat to send an agent to work.

Tracking progress

The Active tasks section shows what is running. The Reports section collects the results. A run has one of these statuses: “Running,” “Waiting for you,” “Complete,” “Canceled,” “Failed,” or “Interrupted.” If an agent asks a question, it waits for your answer.

Model and subagents

The model used by agents is selected through Model for agents & cron jobs on the agent server's page under Local.

The assistant can also delegate a subtask to a subagent without you creating an agent first. The limits are on the Agents page under Subagents.

Cron jobs

Right sidebar → Cron jobs shows a calendar with Day and Month views. New cron job asks for three details:

Type	Agent, Script, or Reminder
Repeat	Once, Daily, Weekly, Monthly, or Custom (using a cron expression)
Time	when to start

You can also simply tell the assistant, for example: “Remind me to check the mail every morning at eight.”

Run now runs a cron job immediately. The Recent runs section shows the results. If a cron job has a goal, only the run that reaches it sends a notification. If a cron job fails several times in a row, it pauses until you enable it again.

Chapter 10 / 16

Images and video

The Images icon at the bottom of the left sidebar opens the area where you generate images and videos. Generation runs on your computer.

The image studio with canvas and controls — Inside Exe AIEnlarge image ↗

What you need

An image model from the catalog's Image tab. You also need the image generation software: under Local → Image server, use Download framework to download it once.

How it works

Write what you want to generate in the prompt bar. Enhance rewrites your prompt in English with more detail; Undo restores your text. Generate starts generation; Stop cancels it.

Setting	Purpose
Model, format	Which model is used and the image's aspect ratio.
Dimensions & generation	Width, height, steps, CFG, and seed.
Style, LoRA, embeddings	Add-ons that guide the image's style.
Input image	An existing image to edit.
Mask	Use the brush to mark the areas to regenerate.
Reference images	Images that guide the model during editing. Some models accept multiple reference images.
Face	ADetailer finds faces in the finished image and sharpens them.
Face transfer	PuLID transfers a face from a photo.
Pose & structure	ControlNet guides the pose or outlines.
Upscaling	Hires fix generates a larger version of the image in a second pass.

Switch between Image and Video. Video has four modes: “Text to video,” “Image to video,” “First and last frame,” and “Video to video.” Video is still experimental. The available options depend on the selected model.

Finished images appear in the Gallery in the right sidebar. The assistant can also generate images in the chat when you ask. The Remove image metadata switch under Settings → Memory removes the prompt, settings, and EXIF data from each image.

Chapter 11 / 16

Dictation and conversation

Both require the speech server under Local and a speech model from the catalog's Speech tab. Speech recognition and speech generation run on your computer.

Dictation

Click the microphone to start dictation; click again to stop. A preliminary transcript appears as you speak.

Optionally, a model can polish the text. Select the model on the speech server's page under Cleaned up by. With Automatic dictation, the text is sent automatically when you finish; otherwise, you send it yourself.

Conversation

Hold the microphone button for one second. The app listens, thinks, and speaks. Enable Read responses aloud and choose a Voice to hear every response. The voice uses a separate model from the catalog.

The conversation ends automatically if you do not respond for three rounds.

The shortcuts ⌘ ⇧ M and Ctrl + Shift + M work system-wide: they bring the window to the front and do the same thing as clicking the microphone, even when another app is in the foreground.

Chapter 12 / 16

What the assistant knows

Four things accompany every conversation. You can read, change, and disable each of them.

Instructions

Settings → Instructions. This is where you describe how the assistant should behave. Suspend instructions for this chat disables them for the open chat.

Memory

Settings → Memory. Rules and facts that apply across conversations. The assistant adds entries itself; you can read and edit them, or remove them all with Clear.

Send to local models — memory entries stay on this computer. Send to cloud models — memory entries are sent to the provider with every request.

Your name

You will find it under Preferred name on the Instructions page. The Greeting & preferred name switch displays the greeting and includes your name in requests to models. Turning it off disables both. With cloud models, your name is sent to the provider.

Skills

A skill is a written workflow. Invoke it by typing a slash followed by its name. The included skills are /cleanup, /cronjob, /grill-me, /handoff, /hilfe and /image-generator.

Settings → Skills: Suggest skills automatically lets the model select a skill without you naming one. Write a skill creates a custom skill. Reset restores a modified built-in skill to its original version.

The notes in exe.md in each approved folder are included as well (see Folders and documents).

Three limits on the Instructions page

Tool rounds	How many times the model may use a tool within one response. Tasks involving many files need more. Auto means 50.
Command timeout	How long a shell command may run, in seconds. Downloads and builds need more time. Auto means 120.
Context compression	The context usage threshold at which a chat summarizes itself, as a percentage. The range is 1 to 85; Auto means 80.

Chapter 13 / 16

Settings

Open Settings using the Settings icon at the bottom of the left sidebar, or press ⌘ , on macOS or Ctrl + , on Windows.

Group	Page	What you will find
General	Language	Auto (device), English, or Deutsch. Changes take effect immediately.
	Appearance	Theme, message color, code colors, and layout mode.
	Notifications	Sounds, volume, popup duration, and colored indicators for agents, cron jobs, the terminal, and chats.
	Keyboard shortcuts	All shortcuts at a glance.
Assistant behavior	Memory	What it remembers and where that information is sent.
	Skills	Built-in and custom skills.
	Cron jobs	All cron jobs, their calendar, and their scripts.
Prompts & agents	Instructions	Your instructions, preferred name, and the three limits.
	Agents	Create and edit agents; configure subagents.
Connections & keys	MCP servers	Connect apps.
	Tools	Enable or disable individual tools.
	Credentials	For example, your Hugging Face token.
Files	Configuration, tools file	For anyone who prefers to edit files directly. Changes take effect the next time the app starts.
Help		This manual.

Appearance

Choose from five themes — Pearl, Exe, Dark Matter, Heaven, and Moune — plus Custom gradient with your own colors and Auto, which follows your device's settings. A theme changes the entire window: frame, lines, messages, and code blocks.

Message color sets the color of your own messages. Under Code colors, choose from these color schemes: Exe, Tokyo Night, One Light, GitHub, GitHub Dark, Dracula, Nord, Solarized Light, and Solarized Dark. Layout mode adjusts the font size, chat width, and spacing between messages.

Updates

At startup, the app checks for a newer version. If one is available, the New version available card appears with Download, Install and Later.

Use Check for updates in Settings to check immediately.

Chapter 14 / 16

Keyboard shortcuts

The full list is under Settings → Keyboard shortcuts.

Action	macOS	Windows
New chat	`⌘ N`	`Ctrl + N`
Search chats	`⌘ K`	`Ctrl + K`
Toggle the left sidebar	`⌘ B`	`Ctrl + B`
Toggle the right sidebar	`⌘ J`	`Ctrl + J`
Settings	`⌘ ,`	`Ctrl + ,`
Microphone, system-wide	`⌘ ⇧ M`	`Ctrl + Shift + M`
Send	`Enter`
New line	`Shift + Enter`
Stop response	`Escape`

Chapter 15 / 16

Your data

Exe AI is designed to keep your data with you.

The app is accessible only from your own computer (address 127.0.0.1). Chats, settings, keys, and models are stored locally in files and a database.
With a local model, nothing you type leaves your computer. With a cloud model, your text is sent to the provider.
API keys are stored in a local file with restricted permissions and are never displayed again. The interface only receives information about whether a key has been set.
Your memory entries are sent to cloud providers only if you enable that option. Your name accompanies every conversation, including those with cloud providers, as long as “Greeting & preferred name” is enabled.
Models are downloaded directly from huggingface.co.

System	App data folder
macOS	`~/Library/Application Support/Exe AI Terminal`
Windows	`%APPDATA%\Exe AI Terminal`

Chapter 16 / 16

Troubleshooting

Common messages and what to do about them.

Message	What to do
No model available	No server is running and no provider is connected. Start the model server under Local or connect a provider under Cloud.
No model in the folder yet	First, download one from the Catalog.
This model cannot call tools	Choose a model marked “Calls tools” in the catalog.
Document too large	The document does not fit entirely in the context. It remains in the chat's folder, and the assistant reads individual passages from it.
Context window is full	Start a new chat or set a larger context on the model's page.
The speech server is not running	Start it under Local using the Speech server card.
Microphone access is blocked	Grant access in your computer's system settings.
Missing key	Under Cloud, open the provider and enter your key.

You can also ask the assistant directly. Use /hilfe to ask questions about the app.

Everything Exe AI can do — and where to find it.

What do you need help with?

Search results