AI
Free & PremiumThe AI page runs large language models right on your Mac. It uses Ollama, a local server for language models that WebDonkey manages as a service. With a local model there are no token costs, and your chats, files and code stay on the Mac. The exception are Ollama's cloud models, which you add yourself: they send everything to ollama.com.
What you do here
- Chat with a local model and attach files as context: Chats
- Set model, memory and specialist once and start new chats from it: Configurations
- Create reusable experts with instructions and files (Premium): Specialists
- Run coding agents such as Aider or Qwen Code on the local model in terminals (Premium): CLI
- Download models, rated for your Mac: Models
- Add any model from the Ollama library or Hugging Face: More models
Page layout
At the top right you switch between the tabs Chats, Specialists, CLI and Models. The page remembers the tab you had open until you quit the app. ⌘K jumps straight into a tab, for example with AI › CLI.
If Ollama is not running, the page shows Ollama is not running instead of the tabs, with two buttons:
- Start Ollama starts the service. The tabs then appear on their own, the page checks every few seconds.
- Open Services opens the Services page. If it says "Ollama does not appear to be installed yet.", install Ollama there first.
If your Mac already runs its own Ollama app, WebDonkey uses it as long as it answers on the default port.
Memory and context
A language model needs memory for two things: for itself and for the context window, the amount of text (in tokens) it can keep in view during a conversation at once. In WebDonkey you set a single value for both, the memory (in the app: Total memory budget for the AI (model + context)). Whatever is left after the model becomes the context window.
| Choice | Meaning |
|---|---|
| Auto · … GB | depends on your Mac's memory: 4 GB below 16 GB of RAM, 6 GB from 16 GB, 8 GB from 32 GB, 16 GB from 64 GB, 24 GB from 128 GB |
| 4, 6, 8, 12, 16, 24, 32 GB | fixed values |
| Custom… | any value from 1 GB, in GB or MB. A smaller or half-typed value snaps back to the valid one when you leave the field |
On hover the selector shows how many tokens the value buys for the selected model, such as "8 GB ≈ 33k Tokens". If the model alone exceeds the budget, it says "model alone exceeds this budget".
You choose the memory in several places, and each applies only to itself: in every chat, in every configuration, in the bar of the CLI tab and in the header of the Models tab. That last one sets the model and memory new chats start with.
Cloud models
Some Ollama models do not run on your Mac but on Ollama's servers. You recognise them by (cloud) after the name and by the Cloud badge. When one is selected, a yellow notice appears: your prompts, attached files and the answers go through ollama.com and leave the Mac. To keep everything local, pick a different model.
What happens in the background
- WebDonkey stores chats, configurations and specialists in its database on your Mac.
- Models live in Ollama's standard folder,
~/.ollama. An existing Ollama installation sees the same models, nothing is stored twice. - The Ollama that WebDonkey starts only listens on
127.0.0.1and cannot be reached from the network.
Free and Premium
| Feature | Free | Premium |
|---|---|---|
| Chats, attaching files, memory per chat | yes | yes |
| Configurations for chats | yes, without specialist | yes, with specialist |
| Downloading and managing models | yes | yes |
| Specialists | no | yes |
| CLI (coding agents) | no | yes |
If you switched on Hide Premium features under Settings › Navigation (see Navigation), the Specialists and CLI tabs disappear on Free.
Command line
wd ai lists your chats, wd ai configs the configurations, wd ai models the installed models. More under More tools.
Chats Free
In the Chats tab you hold saved conversations with a language model. Every chat has its own settings: its model, its memory and, with Premium, its specialist. One chat can run a code question on a coder model while another uses a small, fast model, without you switching back and forth. Chatting is free.
Where to find it
AI, tab Chats. Or press ⌘K and type AI › Chats.
How to do it
- Click New chat at the top left. The chat starts with the model and memory from the Models tab.
- If needed, pick another model or memory in the chat header.
- Type your message at the bottom. Enter sends, Shift+Enter adds a new line.
- The answer streams in word by word. Until the first word arrives, a small donkey shows that the model is loading and thinking. The stop button cancels, whatever already arrived is kept.
Your first message becomes the chat's title, cut to 60 characters.
The list on the left
| Element | Meaning |
|---|---|
| New chat | creates an empty chat |
| Configurations | saved combinations of model, memory and specialist. A click starts a new chat from one, see Configurations |
| Chats | your conversations, the most recently used first. A cap icon marks a chat with a specialist |
| Time under the title | when you last used the chat, such as "5 minutes ago". On hover you see the exact date |
| Pencil (Rename chat) | renames the chat. Double-clicking the title does the same |
| Bin | deletes the chat for good after the question Delete chat? |
While renaming, Enter or a click elsewhere saves the new name (at most 80 characters). Esc cancels, an empty name keeps the old one.
The chat header
| Element | Meaning |
|---|---|
| Context usage | bar and figure such as ≈1.2k / 33k: how full the model's context window is in this chat. Before the first answer this is an estimate (≈), afterwards the real value from Ollama. From 80 % the bar turns yellow |
| Model | this chat's model. Cloud models carry (cloud) |
| Memory | this chat's memory budget, see AI |
| Specialist | only with Premium and when specialists exist: binds a specialist to the chat. No specialist removes it |
| Bookmark icon (Save as configuration) | saves this chat's model, memory and specialist as a configuration |
What you change here applies to this chat only. The other chats and the defaults in the Models tab keep their values. If a chat's model gets deleted, the chat uses the model from the Models tab.
Files as context
- Click the paper clip left of the input field (Attach files or folders as context).
- Pick one or more files or folders.
- The attachments show as small chips above the input field. The X on a chip removes it.
WebDonkey reads the attachments again with every message, so changed files reach the model in their current state. So that a large folder does not blow up the window, limits apply: at most 40 files, 48 KB per file, 256 KB in total. WebDonkey skips binary files and hidden folders, as well as .git, node_modules, vendor, target, dist, build, .venv and __pycache__.
If the context window gets tight, more memory in the chat header or a model with a larger window helps.
Answers
Answers are formatted, with lists, headings and code blocks. Links open in your browser, never in the app window. The chat does not load images from the web on its own, instead you see Image not loaded. A click opens the image in the browser. That way no outside server learns when you read the answer.
What happens in the background
Messages go to the local Ollama, together with the context window your memory buys. When a chat changes its model or window, Ollama reloads the model with the next message, which takes a moment. Your message is saved right away, the answer once it is complete. If you leave the page while an answer is streaming, it stops and the part that already arrived stays in the chat. Half-typed text in the input field and the attachments are kept until you quit the app.
Free and Premium
Chatting, attachments and the per-chat settings are free. Only binding a specialist needs Premium. If Premium ends, the binding stays saved but has no effect: the chat then answers without the specialist's instructions.
Troubleshooting
| What you see | What to do |
|---|---|
| "Install a model in the Models tab first." | download a model in the Models tab |
| the first answer takes long | normal the first time a model loads, it gets faster afterwards |
| the model forgets the start of the conversation | the context window is full: raise the memory or start a new chat |
| yellow cloud notice above the chat | the model runs in Ollama's cloud, pick another one for local work |
Command line
wd ai chats lists the chats with model and last change. wd ai rename "New chat" Shop checkout renames a chat. More in More tools.
See also
Configurations Free & Premium
A configuration saves a chat's settings under a name: the model, the memory and, with Premium, a specialist. Clicking it starts a new chat with exactly those values. That pays off as soon as you have recurring tasks, such as "Laravel with 16 GB" for code questions or a small fast model for short texts.
Where to find it
AI, tab Chats, the Configurations section on the left, between New chat and the list of chats.
How to do it
Create a new configuration
- Click the plus next to Configurations (New configuration). The dialog takes the model and memory from the Models tab.
- Enter a name, for example "Laravel with 16 GB".
- Choose Model, Memory and, with Premium, a Specialist.
- Click Save.
Save one from a chat
- Open the chat whose settings you want to keep.
- Click the bookmark icon in the chat header (Save as configuration).
- The dialog takes this chat's model, memory and specialist. Enter a name and save.
Start a chat from a configuration
Click the configuration in the list (Start a new chat with this configuration). The new chat appears at the top of the list and opens right away. Whatever you change in the chat afterwards stays in that chat, the configuration does not change.
Fields and options
| Field | Meaning |
|---|---|
| Name | how the configuration is labelled in the list. Required |
| Model | one of your installed models. Cloud models carry (cloud) |
| Memory | Auto, a fixed value or Custom…, see AI |
| Specialist | only with Premium and when specialists exist. No specialist leaves it out |
Under each name the list shows a summary, such as qwen3:8b · 16 GB · Laravel expert.
Editing and deleting
Hover over a configuration:
- The pencil (Edit configuration) opens the dialog with the saved values.
- The bin asks Delete configuration? and removes only the configuration. Chats you started from it stay.
If a configuration's model has been deleted in the meantime, the new chat starts with the model from the Models tab.
Free and Premium
Configurations with model and memory are free. The specialist part needs Premium. Without Premium the Specialist field is missing, the summary shows no specialist, and a chat from a configuration with a specialist starts without one.
Configurations for the CLI
The CLI tab has configurations of its own. They store CLI, model, memory and folder for the coding agents and do not appear in this list.
Command line
wd ai configs shows the configurations for chats and for the CLI. More in More tools.
See also
Specialists Premium
A specialist is a reusable expert: fixed instructions about who it is and how it should answer, plus files or folders as context if you like. Bind it to a chat, and the instructions and files go along with every answer. You no longer have to explain to a model each time that it should answer as a Laravel expert and know the project docs. Specialists are a Premium feature, the plain chat stays free.
Where to find it
AI, tab Specialists. Or press ⌘K and type AI › Specialists. The tab is for managing them, you ask your questions in a chat.
How to do it
- Click New specialist at the top left.
- Enter a name, such as "Laravel expert" or "SEO reviewer".
- Write the instructions: who is this specialist, how should it answer, what should it pay attention to?
- Under Context files & folders, attach material with Add file or Add folder, such as project docs or a folder with conventions.
- Click Save. Name and instructions are required.
- Click Start chat on the right. WebDonkey creates a saved chat bound to the specialist and switches to the Chats tab.
You can also bind a specialist in the header of any chat with the Specialist selector, or in a configuration.
The view
| Element | Meaning |
|---|---|
| List on the left | all specialists. On hover, a pencil (edit) and a bin appear |
| Edit | opens the Edit specialist dialog |
| Start chat | starts a saved chat with this specialist, using model and memory from the Models tab |
| Instructions | the text placed in front of every conversation |
| Context files & folders | the attached paths. Without paths it reads "No context files" and only the instructions apply |
Deleting asks Delete specialist? and removes only the preset. Your files stay untouched. Chats that were bound to it carry on without a specialist.
What happens in the background
In a chat, WebDonkey sends the instructions and the content of the context files ahead of the conversation. It reads the files at the first question after you open the chat and reuses them for the following questions. When you reopen the chat or switch the specialist, it reads them fresh.
The chat header shows a line for this: before the first question something like "2 context paths", afterwards "14 files in context", with truncated added for large folders. The same limits as for attachments apply: at most 40 files, 48 KB per file, 256 KB in total. WebDonkey skips binary files and folders such as .git or node_modules.
Free and Premium
Without Premium the tab shows the lock Specialists are a Premium feature. With Hide Premium features it disappears completely. If Premium ends, your specialists stay saved. Bound chats then answer without them, and you can remove the binding in the chat at any time with No specialist.
See also
CLI Premium
In the CLI tab you start terminal coding agents on your local model. A coding agent reads your code, proposes changes and runs commands, much like Claude Code, but with a model that runs on your Mac. WebDonkey wires the agent up to the local Ollama for you and opens it in a terminal on the page, as many side by side as you like. This is a Premium feature.
Where to find it
AI, tab CLI. Or press ⌘K and type AI › CLI.
How to do it
- In the bar at the top, choose the Model, for example
qwen3:8b. - Choose the CLI, the agent program: Aider, Qwen Code, Codex CLI (--oss), Goose, OpenCode or Crush. If the program is missing, the right side shows Install …, one click installs it.
- Set the memory. It decides the agent's context window.
- Click the folder button and choose the Working directory, usually a site's folder.
- Click Create terminal. The agent starts in a terminal below the bar.
Every further click on Create terminal opens an additional terminal with the current values. Aider and Goose can work in two projects at the same time, for example.
The bar
| Element | Meaning |
|---|---|
| Configuration selector | New configuration or a saved one. A saved one fills in model, CLI, memory and folder |
| Save | without a selected configuration, WebDonkey asks for a name and creates a new one. With a selected configuration the button only becomes active once you change something, and then writes the change into it |
| Bin | deletes the selected configuration. Running terminals keep going |
| Model | the agent's model, independent of your chats |
| CLI | the agent program. The number in brackets is the context the CLI itself occupies before your first message (see below) |
| Memory | the memory budget, see AI |
| Folder button | shows the end of the working directory and opens the picker |
| In the editor | only when WebDonkey Editor is installed: opens the same setup as a terminal window in the editor |
| Mac Terminal | opens the same setup in the macOS Terminal |
| Create terminal | starts the agent in a new terminal right here on the page |
Below the bar, the left side shows the full path of the working directory, a click opens the folder. The right side shows the context arithmetic. With two or more terminals you choose Side by side or Stacked there.
Working directory
The folder button opens the Working directory dialog:
- Sites: your sites with their project folder, a click picks it.
- Saved folders: your own folders that stay in the list permanently. The X only removes a folder from the list, it stays on disk.
- Add folder … picks an existing folder and remembers it.
- Create new … first asks for the parent folder and then for the name. The name may not contain slashes or start with a dot.
Without a choice of your own, the agent starts in the folder of your first site.
Context arithmetic
Every CLI sends the model its own instructions and the description of its tools before your first question. That space is missing from the context window:
| CLI | occupies about |
|---|---|
| Aider | 4k tokens |
| Qwen Code | 35k tokens |
| Codex CLI (--oss) | ≈10k tokens |
| Goose | ≈8k tokens |
| OpenCode | ≈12k tokens |
| Crush | ≈10k tokens |
Values with ≈ are estimates, Aider and Qwen Code report their own cost. The line at the right below the bar does the maths for you: context window minus CLI gives the space left free for you, plus the memory in GB. On hover it shows the full calculation with model size and cost per token. While a terminal runs, it also shows model loaded. If the running model still has a different window than your current setting, restart session for …k appears: end the terminal and create a new one.
If the window is smaller than one and a half times what the CLI occupies, a yellow notice warns you. More memory helps then, or a model with a cheaper context, or Aider, which only takes a few thousand tokens.
Models without tool calls
Agents work by having the model emit structured tool calls, such as "read this file". Some models cannot do that reliably. In the catalog they carry the label no tools. If you pick such a model with any CLI other than Aider, a notice warns you: the agent then only talks about actions instead of doing them. Use such models with Aider, which works with diffs, or pick an agent-ready model such as qwen3 or gpt-oss.
The terminals
Each terminal shows CLI, model, folder and memory at the top. The X (End terminal) ends the agent. When you switch tabs or pages, the agents keep running, and the terminal is back when you return.
The CLIs keep their own confirmation prompts before they change files or run commands. WebDonkey never switches on a mode that allows everything without asking.
If you start a terminal while others with a different context window are running, the new terminal shows a notice: Ollama is not restarted for it, so the new CLI keeps the window already set until the others are ended. Only Aider sends its own window.
What happens in the background
- Every CLI is connected to the local Ollama at
127.0.0.1:11434. With a local model your code stays on this Mac. - WebDonkey switches off the CLI vendors' usage statistics at startup, using the switches each CLI offers for this.
- So that the context window is right, Aider gets a small settings file per model and window in WebDonkey's data folder, and Codex gets the value as a start option. For the other CLIs WebDonkey sets the default window of its Ollama and briefly restarts the service when the value changes and no other CLI terminal is running. A separate Ollama app outside WebDonkey keeps its own settings.
- OpenCode and Crush only find local models through their config file. WebDonkey adds a provider "Ollama (lokal)" with your installed models there and sets the selected one as default: in
~/.config/opencode/opencode.jsonand~/.config/crush/crush.json. Before that it puts a copy with the extension.webdonkey-backupnext to it. A file with comments is left unchanged. - Saved configurations live in WebDonkey's database.
Free and Premium
The whole tab needs Premium. Without Premium it shows the lock The agent is a Premium feature, with Hide Premium features it disappears completely. Saved configurations are kept and can be deleted on the command line even without Premium.
Troubleshooting
| What you see | Why | What to do |
|---|---|---|
| the agent does not see your question or answers cut off | the context window is almost entirely taken by the CLI | raise the memory or use Aider |
| the agent only describes changes | the model emits no tool calls | pick an agent-ready model |
| OpenCode or Crush only show cloud providers | the config file is not plain JSON (with comments, for example) or Ollama was not running at the start | add the provider by hand, or start Ollama and create the terminal again |
| Create terminal is greyed out | a model or a working directory is missing | download a model in the Models tab and choose a folder |
Command line
wd ai configs shows the configurations. wd ai cli-add Shop --model qwen3:8b --folder ~/Sites/shop creates one (options --cli and --memory), wd ai cli-delete deletes one after asking. More under More tools.
See also
Models Free
In the Models tab you download and manage the language models. The catalog rates every model for your Mac: what fits comfortably into memory is at the top. This is also where you set the model and memory new chats start with.
Where to find it
AI, tab Models. Or press ⌘K and type AI › Models.
How to do it
- Look at the top for what your Mac offers: chip, unified memory and cores.
- Find a model rated Recommended in the Catalog.
- Click Download. A bar shows the progress.
- When the model is ready, it appears under Installed models, and the bell tells you.
- Pick it in the page header as the model for new chats.
Page header
Only in this tab there are two selectors at the top right:
- Model: the model New chat and Start chat on a specialist begin with.
- Memory: the memory budget for new chats. The catalog also calculates the context windows from it, see AI.
Existing chats keep their own values.
Your Mac
The card at the top names chip, unified memory and cores. On Apple Silicon models run with graphics acceleration in unified memory. On an Intel Mac only the processor computes, which is much slower: stick to small models there.
Installed models
Installed models lists everything Ollama has on disk, including models you downloaded with the Ollama CLI. Where known, each model shows the labels agent-ready or no tools, the parameter size, the quantisation (such as Q4_K_M) and the size on disk. The bin asks Delete model? and removes the file. You can download it again at any time.
Catalog
The Catalog is rated for this Mac and sorted by what runs best here.
| Rating | Meaning |
|---|---|
| Recommended | fits comfortably into unified memory, next to your other apps |
| Tight fit | runs, but only with a small context |
| Slow on Intel | runs on an Intel Mac but answers much more slowly. On Intel this is the best rating |
| Checking… | WebDonkey is still reading the hardware |
| Too big for this Mac | will not load, Download is disabled |
More details per model:
- agent-ready or no tools: whether the model emits reliable tool calls. Only agent-ready models work with the coding agents in the CLI tab, models with no tools only in chat and with Aider.
- Keywords such as Chat, Code, Vision, Reasoning and Fast.
- A short description. Models that only complete code carry the note "No chat". They are meant for the suggestions while typing in WebDonkey Editor.
- The context window at your memory, such as
≈32k Ctx. For installed models the value is exact, otherwise estimated (≈). - The download size.
Downloading
Download starts the download. A progress bar with a percentage and a stop button for cancelling replaces the button. The download keeps going when you switch pages. When it is done, the bell reports that the model is ready, and the catalog shows Installed.
Ollama downloads in chunks. If a download is interrupted, for example because you quit the app, a new click on Download picks up where it stopped.
What happens in the background
Models live in Ollama's standard folder, ~/.ollama, and are shared with an existing Ollama installation. A model is often several gigabytes. You delete it here or with Ollama itself.
See also
More models Free
Under More models you add any model from the Ollama library or Hugging Face by name, even if it is not in the catalog. Whatever Ollama does not report about such a model, the context window for example, you fill in yourself. The section sits at the bottom of the Models tab.
Where to find it
AI, tab Models, the More models section at the very bottom.
How to do it
- Click Discover models. The ollama.com search opens in your browser.
- Open a model there and copy its name with the tag, such as
qwen3:8b. - Paste it into the field and click Download.
- A progress bar shows the download, the stop button cancels.
- When the model is ready, it appears in the list below. Use makes it the model for new chats and switches to the Chats tab.
What the field understands
You can paste whatever the page shows:
- the name with the tag, such as
qwen3:8b, - the whole command
ollama run qwen3:8borollama pull qwen3:8b, - the address from the browser, such as
https://ollama.com/library/qwen3:8b, - a GGUF model from Hugging Face as
hf.co/user/repo:quant, such ashf.co/bartowski/…-GGUF:Q4_K_M, or the address of the Hugging Face page. The repository must be public and contain GGUF files. GGUF is the file format Ollama loads models in.
If the name is a cloud model, a notice warns you before the download: everything you send it goes to ollama.com.
The list
Each row shows a model you added:
| Element | Meaning |
|---|---|
| Params, Context | parameter size and context window. Values with the pencil mark ✎ come from your notes, the others from Ollama |
| Cloud | the model runs on Ollama's servers, prompts leave the Mac |
| Quantisation, family, RAM | more details, RAM only from your notes |
| Add missing specs | appears when Ollama does not report parameter size or context window |
| Size | space on disk |
| Use | makes the model the one for new chats and opens the Chats tab |
| Arrow (Update (pull again)) | downloads the latest build of the same tag |
| Notebook (Notes & specs) | opens the notes for the model |
| Bin | deletes the model after confirmation, your notes on it as well |
Notes & specs
Enter what Ollama does not report:
| Field | Meaning |
|---|---|
| Parameter size | such as 8B |
| Context window (tokens) | how much text the model keeps in view, such as 32768. If Ollama reports none, the chat uses this value. Without it Ollama would use a small default and cut long conversations short |
| Recommended RAM (GB) | how much memory the model needs |
| Notes | free text, such as what it is good at, source or licence |
If Ollama already knows a value, that value with the suffix from Ollama stands in place of the field. Ollama's value always wins. Save stores your entries, Cancel closes the notes.
Cloud models
Some Ollama models run in the cloud, their tag ends in cloud, such as gpt-oss:120b-cloud. WebDonkey also recognises them when a local copy under another name points to a remote server. They carry the Cloud badge, and wherever you pick them a yellow notice appears: prompts, attached files and answers go through ollama.com. To keep everything local, pick a different model.
Troubleshooting
| What you see | What to do |
|---|---|
| "That is not a valid model name." | check the spelling, something like qwen3:8b or hf.co/user/repo:Q4_K_M is expected |
| "… was not found." | check name and tag on ollama.com. On Hugging Face the repository must be public and contain GGUF files |
| "Could not download …: no connection to the model registry." | check your internet connection and try again |
The X next to an error message closes it.