Buy

AI

Free & Premium

The AI window (⇧⌘I) brings local language models into the editor. You download models, chat about your code, have code explained or adjusted from the context menu, get grey suggestions while typing and start a coding agent on a local model. The models run on your Mac: Ollama, LM Studio, llama.cpp, Jan, MLX or a server of your own with an OpenAI-compatible interface. The editor never accepts an internet address as an AI server.

Two exceptions are your deliberate choice: models marked Cloud run on ollama.com, and AI: Explain and AI: Adjust can work with Claude Code instead of the local AI.

What you do here

Layout of the window

Area What is there
Header Providers, on the right Start chat (a new chat with the selected model) and Look for local AI (↻).
Providers One block per provider with a dot (green = running), name, state and number of models, its models below. Under Ollama's models sits Download a model. Providers that are not installed share one line.
Configurations Saved templates for new chats.
Specialists Roles with instructions and context files (Premium).
Chats The last five chats, plus All chats.
Context The memory budget and the context window it buys for the selected model.
CLI Which agent CLI Agent starts (Premium).

WebDonkey keeps configurations, specialists and chats in its database. Without WebDonkey the editor uses its own. The window only looks for servers when it opens and when you click ↻, never in the background.

Free and Premium

Free: the window, all providers and models, downloads, the memory budget, the chat, chats, configurations without a specialist, and AI: Explain and AI: Adjust. Premium: Specialists, Suggestions while typing and the Agent with a local model. Prices are on webdonkey.io/pricing.

Privacy

Files that look like they hold secrets never go to the AI, not in the chat, not for suggestions while typing and not for AI: Explain and AI: Adjust. The patterns are under Settings › AI › Never send as context. The defaults include .env, .env.*, *.pem, *.key, id_rsa*, auth.json, .npmrc and credentials*. If you don't want the AI at all, turn AI off under Settings › Features › Tool windows. Then there is no window, no chat, no suggestions and no context menu entries.

Providers and models Free

At the top of the AI window are the providers, the programs on your Mac that run language models, and below them their models. Here you see what is running and what a model can do, and you pick the model the chat and the editor work with.

Where to find it

Open the AI window with ⇧⌘I or its icon in the right stripe. When it opens, it looks for running servers, the line "Looking…" says so. Look for local AI (↻) looks again, for example after you started a server.

Providers

Provider Address How to start it
Ollama 127.0.0.1:11434 Runs as a WebDonkey service, see below. Without WebDonkey, the Ollama app or ollama serve.
LM Studio 127.0.0.1:1234 Start the server in LM Studio (Developer tab) or lms server start.
llama.cpp 127.0.0.1:8080 llama-server -m <model.gguf>
Jan 127.0.0.1:1337 Turn on the local API server in Jan's settings.
MLX 127.0.0.1:8080 mlx_lm.server --model <model>
Own URL your address Set it under Settings › AI › Own server URL.

Ollama is always in the window, your own URL as soon as you set one. The other providers show up when they are running or installed. Whatever is neither running nor installed shares one line ending in not installed. That line's tooltip tells you how to start each server.

Each provider shows a dot (green = running), the state running with its version, off or not installed, and the number of models. If an installed server is off, a hint below tells you how to start it. If it runs without a model, it says "Running, but no model loaded."

When Ollama is not running

The window then shows the box Ollama is not running:

  • WebDonkey is running: "Chat, agents and ghost text use the local Ollama server." Start Ollama starts WebDonkey's Ollama service, then the window keeps looking for a few seconds until Ollama answers. Open AI page in WebDonkey takes you to the models in WebDonkey.
  • WebDonkey is installed but not running: "WebDonkey runs Ollama for you. Start WebDonkey first, then Ollama." with Start WebDonkey.
  • Without WebDonkey: "Start Ollama yourself (the Ollama app or ollama serve), then look again."
  • If Ollama is not installed on this Mac at all, Install in WebDonkey shows up as well. It opens WebDonkey on its AI page at Ollama, where you install it.

Models

Each model row shows:

  • the model's name, for cloud models the mark Cloud,
  • size (or parameter count), context length (for example "33K context") and quantisation,
  • chips for what the model can do: chat, code, vision, completion (for suggestions while typing), tools, thinking (reasoning) and embedding.

Clicking a model with chat selects it, the row is then highlighted. That model is what Start chat uses, what configurations suggest, and what AI: Explain and AI: Adjust use while The AI window's model is set there. If you picked none, or the picked one is not available right now, the editor takes the first local chat model. The editor remembers your pick.

Below each model are buttons:

Button What it does
Chat Opens a new chat with this model.
Terminal Starts the model in the terminal when the provider has a command for it: ollama run, lms chat, llama-cli -m or mlx_lm.chat --model. Jan and your own URL have none.
Agent Only for Ollama models without the Cloud mark: starts a coding CLI with this model in the project folder (Premium, see Agent with a local model).
Trash can Ollama only: deletes the model, see Downloading models and memory.

Models without chat ability, pure embedding models for example, cannot be picked for the chat.

Cloud models

Models marked Cloud run on ollama.com, not on your Mac. The tooltip says so: "Runs on ollama.com: prompts leave this Mac." The editor never picks them on its own, only when you click them.

Own URL and AI on another computer

Under Settings › AI, Own server URL takes a server of your own that speaks the OpenAI API, for example http://127.0.0.1:5000. Only this Mac is allowed. If the address should point to another computer in your network, turn on Allow LAN hosts. The first time, the editor asks: AI on another computer says that your code, selections and chat then travel over the network to that computer. Use it confirms this for that computer. Only use computers you trust. Internet addresses are never accepted.

Troubleshooting

What you see Why What to do
"Not reachable." at the provider The server does not answer Start the server, then ↻.
"The URL is not valid (http://host:port)." Your own URL is invalid Correct it under Settings › AI.
"That is a computer in your network. Allow LAN hosts in Settings › AI first." The URL does not point to this Mac Turn on Allow LAN hosts.
"Only local AI: an internet address is never used." The URL is a public address Use a local server.
"This provider has no terminal command." Jan and own URLs have no CLI Use the chat.
"llama-cli needs the model file; start llama-server without --alias." llama-server runs with an alias Start llama-server without --alias.
No model can be picked No running provider has a chat model Download a model, see Downloading models and memory.

See also

Downloading models and memory Free

While Ollama runs, you download models right in the AI window and delete them again. A catalog suggests models and tells you whether they fit your Mac. With the memory budget you decide how much RAM a model may take. From it the editor works out the context window, that is how many tokens the model keeps in view at once: your conversation, attached files and its answer.

Where to find it

  • Download a model sits in the AI window (⇧⌘I) below Ollama's models, as long as Ollama runs.
  • The Context section is at the bottom of the AI window.
  • The same memory budget is under Settings › AI › Memory budget (model + context).

Download a model

  1. Below the Ollama models click Download a model. A field and the catalog unfold.
  2. Either type a model's Ollama name into the field, for example qwen3:8b, and click Download (or press ↩). Or click the download icon next to a model in the catalog.
  3. The download shows up as a row with name, bar and percentage. The square Cancel download stops it.
  4. Once the model is there, the editor reports "… is downloaded." and it appears in Ollama's list.

A download keeps going when you close the AI window. It ends when it is done, when you cancel it or when you quit the editor. Ollama keeps what is already downloaded, a new download resumes from there.

The catalog

The catalog lists models you don't have yet, with name, rating, the context window your memory budget would buy for that model (≈, tooltip "Context window at a … GB memory budget") and the download size. It is sorted by rating and size:

Rating Meaning
Recommended A good fit: the model needs at most half of your RAM.
Tight fit "Runs on this Mac, but only with a small context" (up to about two thirds of your RAM).
Slow on Intel "Runs on this Mac, but Intel Macs compute on the CPU only and answer much more slowly".
Too big Does not fit this Mac. The download button is locked.
Checking… Your Mac's details are not in yet.

The rating uses the download size plus about 20 % for the context. On Intel Macs, every model larger than a 5.5 GB download counts as too big.

Delete a model

  1. Hover over an Ollama model and click the trash can (Delete).
  2. Confirm "Delete …?": "The model file is removed from Ollama. You can download it again anytime."
  3. The editor reports "… deleted." and looks again.

Memory budget and context

The memory budget is the RAM the model may take in total: its weights plus context. Whatever is left after the weights (plus 5 % and 0.5 GB of buffers) becomes context. A bigger budget means more context, but also less memory for other programs.

Choice Meaning
Automatic · … GB Follows your RAM: 4 GB, from 16 GB of RAM 6 GB, from 32 GB 8 GB, from 64 GB 16 GB, from 128 GB 24 GB. The default.
4 GB to 32 GB Fixed steps: 4, 6, 8, 12, 16, 24 and 32 GB.
Custom… A field for your value and the Unit GB or MB. At least 1 GB (1024 MB), otherwise the field snaps back to the value in effect.

Below, the section shows for the selected model:

  • Free: in green, how many tokens of context are left. When it gets tight, the number turns orange.
  • the sum, for example "Window 24k at 8 GB". With Premium it also subtracts what the selected agent CLI takes itself: "Window … at … − Aider 4k = …".
  • the model with its figures: "… model + 0.5 GB buffer", the memory per token ("… KB/token") and the maximum it was trained for ("max …").

Hints appear when something does not fit: "The model alone exceeds this budget." or "The window is only sent to Ollama; other servers use their own." With no model picked yet it says "No model picked yet."

The budget is the default for new chats. Each chat also has its own Memory (see Chat), and a configuration brings its own.

What happens in the background

  • Downloads and deletes go to the local Ollama at 127.0.0.1:11434, without WebDonkey, as long as Ollama runs.
  • The editor sends the computed context window to Ollama with every request. Other servers use their own window.
  • If the editor does not know a model's memory per token, it uses the model's trained maximum, at most 32K. It never goes below 2048 tokens.
  • How much RAM your Mac has and whether it is Apple Silicon or Intel, the editor reads once on the first look.

Free and Premium

Downloading, deleting and the memory budget are free. Only the sum with the agent CLI appears with Premium.

Troubleshooting

What you see Why What to do
Download a model is missing Ollama is not running Start Ollama, see Providers and models.
The download button is grey The model is Too big or already downloading Pick a smaller model.
A red notice with the model's name Ollama refused the download, for example because of a wrong name Check the name on ollama.com.
"The model alone exceeds this budget." The weights are bigger than the budget Raise the budget or take a smaller model.

See also

Chat Free

The chat answers questions about your code with a local model. It opens as a tab in the editor area: your chats on the left, the conversation in the middle, the input at the bottom. You attach the open file, a selection or a problem, and code from an answer goes back into your files with one click.

Where to find it

  • Start chat at the top of the AI window (⇧⌘I): a new chat with the selected model.
  • Chat below a model in the AI window: a new chat with exactly that model.
  • Command palette (⌘K): AI: Open chat.
  • From a configuration or a specialist, or you continue a saved chat.

The tab is called "AI · model". If there is no model yet, it says "No local model yet. Open the AI window (⇧⌘I) and pick one."

Layout of the chat tab

Area What is there
Side bar on the left New chat and every saved chat, see Chats.
Header Show chats/Hide chats, the model select ("model · provider"), Memory, Specialist (Premium, once you have specialists), Save as configuration (bookmark icon) and New chat while the side bar is closed.
Conversation Your questions on the right, the answers in a readable column. Text is set as Markdown, code is coloured like in the editor.
Input The field with the buttons File, Selection, Problem, the token gauge and Send.

An empty chat greets you with "Chat with …" and three suggestions: Explain the open file (attaches the file), How can the marked code be improved? (attaches the selection) and Write me a function that …. A click puts the suggestion into the input, you adjust it and send it.

Ask a question

  1. Type your question into the field "Ask about your code…". ↩ sends, ⇧↩ adds a new line. The field grows with your text.
  2. Attach context if you need it. It comes from the code editor you last typed in:
    • File: the whole file.
    • Selection: the marked text with line numbers.
    • Problem: up to five errors and warnings of the file closest to the cursor, each with two lines of code before and after. Each attachment shows as a chip. A ✂ means "shortened to fit the context". The x removes it.
  3. Click the arrow (Send) or press ↩.
  4. While the answer comes in, a line tells you what is going on: "Loading the model…", "Local AI is thinking…" or "Local AI is working…". The Stop button (square) ends the answer, what came so far stays, followed by "Stopped."

Next to the buttons it says "≈ … of … tokens": an estimate of how full the context window gets with the next send. After an answer Ollama reports the real count. If it is too full, the gauge turns orange: "Larger than the context window: the start of the conversation may be cut." A single attachment takes at most half of what is still free.

Model, memory and specialist

In the header you switch, for this chat:

  • Model: every chat model of the running providers. The switch applies to the AI window as well.
  • Memory: Automatic · … GB or a step from 4 to 32 GB. This sets the chat's context window (see Downloading models and memory).
  • Specialist: No specialist or one of your specialists (Premium).

While an answer runs, the selects are locked. What you change here is saved with the chat. Save as configuration takes the three values into a new configuration.

Use code from the answer

Every code block shows its language at the top and three buttons:

  • Insert at cursor: puts the code at the cursor, or over the selection, in the code editor you were in last. ⌘Z takes it back.
  • Apply as diff: opens the menu Diff with file. At the top is the file the question was about, then open tabs in a matching language, then the other open tabs and Other file in the project … (a file search "Search for a file to diff …"). If no tab is open, the file search starts right away.
  • Copy (icon): puts the code on the clipboard.

The diff opens as a tab "AI proposal ↔ file". You take single changes with the arrows in the gutter or all of them with Take all, then save the file as usual. Where the proposal goes, the editor decides like this: in place of the selection the question was about, otherwise over the whole file when the block looks like a complete version, otherwise at the cursor.

Saved or not

Each chat is saved with its first question, the start of that question becomes its title. From then on every message is added, a stopped answer too, as far as it got. Chats live in WebDonkey's database and therefore also show up on WebDonkey's AI page. Without WebDonkey the editor uses its own database. If the editor has no database at all, for example because WebDonkey is too old, the line below the input says "Without WebDonkey this chat is not saved."

Questions from AI: Explain open a chat tab too, but are not saved as a chat.

What happens in the background

  • Every request goes through the editor straight to the local server of the selected provider, not to the internet (except for cloud models).
  • Sent along are the conversation so far, your attachments, a specialist's instructions and files when one is set, and for Ollama the context window.
  • Closing the chat tab stops an answer that is still running.

Troubleshooting

What you see Why What to do
"No file open in the editor." There is no code editor to attach from Open a file and click into it once.
"Nothing selected." Selection without marked text Mark some code.
"No problems in this file." Problem without errors or warnings Nothing to do.
"… looks like a secrets file and is never sent to the AI." The file matches Never send as context Intended. The patterns are under Settings › AI.
"The local AI returned no answer." The model came back empty Ask again or pick another model.
A red message below the answer The server reported an error Check the provider in the AI window.

See also

Chats Free

Every local AI chat stays saved. You continue it later in the editor, rename it or delete it. It is the same list as on WebDonkey's AI page: a chat you started there continues in the editor, and the other way round. Each chat remembers its model, its memory and its specialist.

Where to find it

  • In the AI window (⇧⌘I) the Chats section with the five most recently active chats. All chats (with more than five, All chats (n)) opens a chat tab with its side bar shown.
  • In a chat tab, the side bar on the left with every chat. Show chats and Hide chats (icon at the top left of the tab) toggle it.
  • Search everywhere (⇧⇧), category AI chats: search by title or model, ↩ opens the chat.

What a row shows

An icon (speech bubble, a graduation cap for chats with a specialist), the title, below it when the chat was last active ("5 minutes ago") and the model. The tooltip gives the exact date ("Last active: …"). The newest are at the top, the chat of the current tab is highlighted. Without a title the model is shown.

Continue a chat

  1. Click a chat in the AI window ("Continue in a chat tab").
  2. If it is open in a tab already, that tab comes to the front. Otherwise a new chat tab opens with the whole conversation.
  3. Just keep writing.

The editor uses the chat's own model if a running provider has it. Otherwise it takes the model selected in the AI window.

In a chat tab's side bar, a click shows the chat in that tab. If it is open in another tab, the editor switches there. If the current tab is still answering, the chat opens in a new tab so the answer keeps going.

Start a new chat

New chat at the top of the side bar (or in the header while the side bar is closed) clears the tab for a new conversation. The new chat is saved with its first question.

Rename

  1. Click the pencil (Rename chat) or double-click the row.
  2. Type the new name into the field (at most 80 characters).
  3. ↩ or clicking elsewhere saves, Esc cancels. An empty name keeps the old one.

Delete

  1. Click the trash can (Delete).
  2. Confirm "Delete chat?": "“title” is deleted with all its messages."

A tab showing the deleted chat starts over.

The side bar

  • Whether it is open, the editor remembers the same way for every project and window.
  • If the tab is narrower than about 680 pixels, the side bar folds away. The icon then lays it over the chat. It closes after a pick, a click beside it or Esc.

What a chat remembers

Value Meaning
Model The model you last wrote with.
Memory Automatic or a fixed size in GB, which sets its context window.
Specialist The bound specialist. Without Premium the chat goes on as a plain chat.

What happens in the background

Chats live in WebDonkey's database, without WebDonkey in the editor's own database. The editor loads the list when it shows, after every answer and when the window gets the focus. It does not poll in the background.

Free and Premium

Chats are free. Only working with a specialist needs Premium.

Troubleshooting

What you see Why What to do
The Chats section is missing The editor has no database right now, for example because WebDonkey is older than the editor needs Update WebDonkey.
"No chats yet. A new chat shows up here with its first question." There is no saved chat yet Start a chat.
The chat opens with another model Its model is not running right now Start the provider or pick another model in the chat.

See also

Configurations Free

A configuration saves under a name what a new chat starts with: model, memory and, with Premium, a specialist. Instead of setting the same three values every time, you click the configuration. They are the same configurations as on WebDonkey's AI page.

Where to find it

  • In the AI window (⇧⌘I) the Configurations section.
  • In a chat tab the bookmark icon Save as configuration.

When there are none yet, the section says: "Save model, memory and specialist once and start new chats from it."

Create a configuration

  1. Click + (New configuration) in the Configurations section. Or use Save as configuration in a chat, then the chat's model, memory and specialist are filled in already.
  2. Enter a name, for example "Laravel with 16 GB".
  3. Model: one of the chat models of your running providers. Cloud models carry "· Cloud".
  4. Memory: Automatic · … GB or a step from 4 to 32 GB.
  5. Specialist (only with Premium and when you have specialists): No specialist or one of your specialists.
  6. Save. Without a name or a model the button stays grey.

"A new chat takes these values over. What you change in the chat afterwards stays in that chat." So a configuration never changes a running chat.

Start a chat from a configuration

Click the configuration's row ("Start a new chat with this configuration"). A new chat tab opens with the configuration's model, memory and specialist. If its model is not running right now, the editor takes the model selected in the AI window.

Below the name it says what the configuration brings, for example "qwen3:8b · 16 GB · Laravel expert".

Edit and delete

  • The pencil (Edit) opens Edit configuration with the same fields.
  • The trash can (Delete) asks "Delete configuration?". Only the configuration is removed, chats you started from it stay.

For AI: Explain and AI: Adjust

Under Settings › AI › AI in the code › Local model you can pick a configuration instead of a model. AI: Explain and AI: Adjust then work with its model, memory and specialist (see AI: Explain and AI: Adjust).

What happens in the background

Configurations live in WebDonkey's database, without WebDonkey in the editor's own database. The editor loads them when the AI window or a chat tab shows them, and after every change.

Free and Premium

Configurations are free. Saving one with a specialist needs Premium. When Premium ends, such a configuration starts a chat without the specialist. Listing and deleting always work.

Troubleshooting

What you see Why What to do
The Configurations section is missing The editor has no database right now Update WebDonkey if it is older than the editor needs.
+ is grey No running provider has a chat model Start a provider or download a model.
"No local model yet. Open the AI window (⇧⌘I) and pick one." No model can be reached when starting Start Ollama or another provider.
The Specialist field is missing No Premium, or no specialist yet Create a specialist first.

See also

Specialists Premium

A specialist is a reusable role for the local model: a name, instructions on how it should answer, and optionally files and folders it gets as context. That way you have, say, a "Laravel expert" who knows your conventions, or an "SEO reviewer" with your checklist. Specialists are part of Premium.

Where to find it

In the AI window (⇧⌘I) the Specialists section. Without Premium it shows a lock and "Specialists are a Premium feature". They are the same specialists as on WebDonkey's AI page.

Create a specialist

  1. Click + (New specialist) in the Specialists section.
  2. Enter a name, for example "Laravel expert" or "SEO reviewer".
  3. Write the instructions: who is this specialist, how should it answer, what should it pay attention to?
  4. Under Context files & folders, use Add file or Add folder to add what it should know. The macOS dialog lets you pick several at once. The x on a line takes it out again.
  5. Save. Name and instructions are required.

When there are none yet, the section says "No specialists yet" and suggests setting one up with instructions and optional files. A row shows the name and, when files are attached, "Files and folders: n".

Chat with a specialist

  • Click the row ("Start a new chat with this specialist"). A chat tab opens with the model selected in the AI window.
  • Or pick it in a running chat at the top under Specialist.
  • Or put it into a configuration.

Chats with a specialist carry a graduation cap in the list of chats.

Edit and delete

  • The pencil (Edit) opens Edit specialist.
  • The trash can (Delete) asks "Delete specialist?". Only the preset is removed, your files are not touched.

What happens in the background

With every question the editor adds the specialist's instructions to the model's instructions and reads its files fresh, so changes apply at once. To keep the context small enough for a small model, limits apply:

Limit Value
Files at most 40
Per file at most 48 KB, the rest is cut off with "[gekürzt]"
In total at most 256 KB
Skipped binary files, hidden folders and the folders .git, node_modules, vendor, target, dist, build, .venv and __pycache__

Specialists live in WebDonkey's database, without WebDonkey in the editor's own database. The files themselves stay where they are.

Free and Premium

Creating, editing and using specialists is part of Premium. When Premium ends, you can still see and delete your specialists. Chats bound to a specialist then go on as plain chats. Prices are on webdonkey.io/pricing.

Troubleshooting

What you see Why What to do
Save is grey Name or instructions are missing Fill in both.
A red message below the answer The specialist's files could not be read Check that the files are still there.
The specialist does not know a file It fell under one of the limits Add specific files instead of large folders.
Clicking the row does nothing No Premium, or no model is running Activate Premium or start a provider.

See also

AI: Explain and AI: Adjust Free

The code editor's context menu holds two AI actions: AI: Explain and AI: Adjust. Both ask you what the AI should do, with a suggestion you are free to change. They work either with the local AI on your Mac or with Claude Code. You choose once, the first time, and later under Settings › AI.

Where to find it

Right-click in the code: both actions are at the top of the menu. They apply to the marked code. Without a selection the editor takes the block under the cursor, that is the line and everything below it that is indented deeper, including the closing bracket. AI: Adjust is missing in read-only files.

How to do it

  1. Mark the code, or put the cursor inside the block.
  2. Right-click, then AI: Explain or AI: Adjust.
  3. A dialog opens with the action as its title and the file and lines below, for example src/main.ts:4-9.
  4. The field "What should the AI explain?" or "How should the AI adjust the code?" holds a suggestion: "Explain what this code does and how it works." or "Simplify this code and add short comments without changing its behaviour." It is selected, so you just type over it. ↩ adds a new line, ⌘↩ sends, Esc cancels.
  5. The first time, the dialog also asks What should the AI run on? (see below).
  6. Click Explain, Adjust or, with Claude Code, Start Claude Code.

The one-time choice

Choice What you set with it
Local AI (Ollama & co. on this Mac) Model: The AI window's model, one of your configurations or a specific model. Memory: Automatic · … GB or a step, only for a single model (a configuration brings its own).
Claude Code (With your subscription or your own API key) Access: Claude sign-in (/login) uses your Claude subscription, or an access of your own from Claude access. The pencil opens Manage Claude access. Keys are kept only in the macOS keychain.

"The editor asks this only once. You can change it any time under Settings › AI." After that the dialog shows a line like "With local AI · qwen3:8b" or "With Claude Code · Claude sign-in (/login)" with Change.

With the local AI

  • AI: Explain opens a chat tab with the code attached and your question asked, the answer appears there. This chat is not saved in the list of chats.
  • AI: Adjust shows the proposal right in the code: the old lines are marked, the new code appears below them. At the top of the box a line says what the AI is doing ("Local AI is working…"), next to it Stop, Apply as diff, Accept and Discard. Accept replaces the code, ⌘Z takes it back. Discard or Esc leaves everything as it was.
  • If the proposal or the old code is longer than 18 lines, the editor opens it straight away as a diff tab "AI proposal ↔ file". There you take it hunk by hunk with the arrows in the gutter or all at once with Take all, and save as usual. Small proposals go into the same tab with Apply as diff.

Very long code is shortened to fit the model's context window. If the configured model is not available right now, the editor says "The chosen model is not available right now, the AI uses … Change it under Settings › AI." If there is no model at all, it opens the AI window.

With Claude Code

The editor starts a new Claude Code session in the project folder, in the terminal or in Mission Control, depending on where your Claude sessions appear. If Claude Code is not installed, the editor offers to set it up. The first message holds the file and lines, the code (up to 12,000 characters, for longer code Claude reads the lines itself) and your instruction. For AI: Explain it adds that Claude must not change any file, for AI: Adjust that it should change only this part of the file. Claude's changes arrive as a diff through the IDE bridge.

Claude Code only starts on your click and sends the selection and instruction to Anthropic. Whether that runs on your subscription or your own API key depends on the access you picked.

Settings

Under Settings › AI › AI in the code:

Setting Meaning Default
Explain and adjust with Ask next time, Local AI or Claude Code. "Ask next time" asks the question again. Ask next time
Local model A model or a configuration. Empty means the AI window's model. The AI window's model
Memory for the local model Applies to a chosen model, a configuration brings its own memory. Automatic
Claude access Claude sign-in or your own API key from Claude access. Claude sign-in

Privacy

Files matching Never send as context never go to any AI. The editor then says "… looks like a secrets file and is never sent to the AI." and does nothing.

Free and Premium

Both actions are free. For Claude Code you need a Claude subscription or your own API key.

Troubleshooting

What you see Why What to do
The actions are missing from the menu AI is turned off under Settings › Features, or the cursor is not in the code Turn AI on, click into the code.
Only AI: Explain is there The file is read-only Open the file writable to change it.
"No local model yet. Open the AI window (⇧⌘I) and pick one." No local provider with a chat model is running Start Ollama or download a model.
"The local AI returned no answer." The model returned nothing Try again or pick another model.
Claude does not start The access in the keychain cannot be read or is missing Check it under Manage Claude access.

See also

Suggestions while typing Premium

While you type, a local model suggests how the code could go on. The suggestion sits in grey behind the cursor, similar to Copilot, and comes from your Mac, not from the internet. You take all of it, word by word or none of it. Suggestions while typing are part of Premium.

Where to find it

  • The status bar shows AI at the bottom right. A click opens the menu with the switches.
  • Command palette (⌘K): Suggestions while typing turns it on and off.
  • Settings › AI › Suggestions while typing for the model and the behaviour.

How to do it

  1. Download a completion model, one with the completion chip, for example qwen2.5-coder:1.5b-base (see Downloading models and memory). If there is none, the status bar menu offers to download the right one for your Mac.
  2. Type in the code. After a short pause the grey suggestion appears.
  3. Tab takes the whole suggestion, ⌘→ the next word, Esc drops it. To get a suggestion right away without waiting for the pause, press ⌥\.

On the first key press the editor looks for the model and loads it into memory. Meanwhile the status bar says "Loading the model…".

The status bar menu

Entry Meaning
Suggestions while typing Turns the feature on or off everywhere.
In this project Turns it off for the open project only.
"Download …" Shows up when there is no completion model yet: downloads the smallest one that fits your Mac.
Load a model in WebDonkey Shows up without a model while WebDonkey runs: opens the models on WebDonkey's AI page.
Show/hide AI Shows or hides the AI window.

The AI entry itself shows the state: struck through when off or off in this project, orange when paused, with a lock without Premium. Its tooltip gives the state ("Suggestions while typing: on", "Suggestions paused: battery below 20 %", "No completion model yet. For this Mac: …"), the model, the median response time, how many answers came too late and how many suggestions you took ("Taken: … of … (… %)"). These numbers are counted on your Mac only.

Settings

Under Settings › AI › Suggestions while typing:

Setting Meaning Default
Suggestions while typing On or off. on
Model A specific model. Empty: the smallest installed completion model that fits this Mac. empty
Wait after typing Asks only after this pause, 100 to 1000 ms. Every key press cancels the pending request. 250 ms
Longest suggestion Upper limit in tokens, 32 to 256. Suggestions stop at an empty line anyway. 96
Pause on low battery Below 20 % on battery no requests go out. on

You hide the status bar entry under Settings › Features › Status bar › AI ghost text.

What happens in the background

  • The editor sends the text before and after the cursor to the model, with the file path and language. Only one request is out at a time.
  • Answers that take longer than 800 ms are dropped, so typing stays smooth.
  • Without a model set in the settings it takes a completion model from Ollama, otherwise one from llama.cpp. It never uses a cloud model for this.
  • There are no suggestions in files over 1 MB, in secrets files (patterns from Never send as context), in the middle of a word, and while you are writing a comment at the end of a line.
  • The editor checks the battery every two minutes while the feature is on.

Free and Premium

Suggestions while typing are part of Premium. Without Premium the AI entry shows a lock and "Suggestions while typing are part of Premium." Prices are on webdonkey.io/pricing.

Troubleshooting

What you see Why What to do
"No completion model yet …" No suitable model is installed Download the suggested model from the status bar menu.
No suggestions appear Ollama is not running, the feature is off for the project, or the battery is low Check the state in the tooltip of AI.
"Could not load the model: …" Ollama could not load the model Check Ollama, pick a smaller model.
Many answers too late The model is too big for smooth suggestions Pick a smaller completion model.

See also

Agent with a local model Premium

A coding agent is a command line program that reads and changes files in the project folder and proposes commands, much like Claude Code, but here with a local Ollama model instead of a cloud model. The editor starts the agent CLI set up for the model you pick, in the terminal. This is part of Premium.

Where to find it

  • Agent below an Ollama model in the AI window (⇧⌘I). Cloud models do not have the button.
  • The CLI section at the bottom of the AI window: which CLI Agent starts.
  • The same choice under Settings › AI › CLI.

How to do it

  1. Open the project the agent should work in. It starts in the project folder.
  2. Pick a CLI in the CLI section. Next to it you see installed or not installed.
  3. If it is missing, click Install in WebDonkey. WebDonkey opens on its AI page at that tool, where you install it. Then click ↻ in the AI window.
  4. Below the Ollama model you want, click Agent (tooltip: "Start … with this model").
  5. A terminal titled "CLI · model", for example "Aider · qwen3:8b", opens in the project folder, and the agent runs in it.

If the selected CLI is not installed when you click Agent, the editor asks: "… is not installed. WebDonkey opens at this tool, install it there. Then click ↻ here." with Install in WebDonkey. Nothing is installed without your confirmation.

The CLIs

The number in brackets is the context the CLI itself takes before your first message (≈ means estimated):

CLI Own context
Aider 4k (default)
Qwen Code 35k
Codex CLI (--oss) ≈10k
Goose ≈8k
OpenCode ≈12k
Crush ≈10k

That share comes off the model's context window. With Premium the Context section therefore shows the sum "Window … at … − CLI … = …". If the window leaves little room (less than one and a half times the CLI's share), it warns: "… alone takes … tokens of context, the … window leaves little room. Raise the memory budget or use Aider." You set the memory budget in the same section (see Downloading models and memory).

What happens in the background

  • The editor starts the CLI with the same launch command as WebDonkey's AI page: connected to the local Ollama, with the computed context window, with usage statistics turned off and never with an option that allows changes without asking. So the CLI asks you itself before it acts.
  • The CLI choice is an editor setting. The default is the same as in WebDonkey (Aider).
  • Installing goes through WebDonkey. The editor installs nothing itself.

Free and Premium

The agent is part of Premium. Without Premium, clicking Agent shows "CLIs set up for the local model are part of Premium.", and the CLI section shows the same with a lock. Prices are on webdonkey.io/pricing.

Troubleshooting

What you see Why What to do
"No file open in the editor." No project is open Open a project.
"… is not installed." The CLI is missing on this Mac Install it with Install in WebDonkey.
"WebDonkey is not installed on this Mac." Without WebDonkey the editor cannot have the CLI installed Install WebDonkey or install the CLI yourself.
"Needs WebDonkey on this Mac." The editor has no database right now Update WebDonkey.
The agent handles tools poorly Some models do not call tools reliably Use another model or Aider.

See also

Cookie preferences

This website uses cookies to ensure the best experience possible. Your choice applies to webdonkey.io and the shop at shop.webdonkey.io. More information...

Technically required

Cookies required for this website to function:

Always on
WebDonkey
WebDonkey
Provider: BuI Hinsche GmbH, Neumarkt 7, 06712 Zeitz, Germany
Purpose: Your consent, colour scheme and language for website and shop, price switches; with wd_debug in the address, the measurement events in the developer console for troubleshooting
Cookies and storage: Cookie wd_consent on .webdonkey.io, shared with shop.webdonkey.io (12 months); cookies wd_theme and wd_lang on .webdonkey.io, shared with shop.webdonkey.io, only when you switch colour scheme or language yourself (365 days); localStorage wd:consent, wd:theme, wd:lang; sessionStorage wd:acc, wd:term, wd:debug (only with wd_debug in the address; until the tab closes)

Statistics

Shows us which pages and products people look at on webdonkey.io and in the shop, where visitors come from and how they get to a purchase (with order number, amount and products, without name or email address). Provider: Google. Transfer to the USA possible (EU-US Data Privacy Framework). Your choice applies to webdonkey.io and the shop together.

Google Tag Manager, Google Analytics 4
Google Tag Manager
Provider: Google Ireland Limited, Gordon House, Barrow Street, Dublin 4, Ireland
Purpose: Loads the services you agreed to, here and in the shop. Sets no cookies of its own.
Cookies and storage: none
Privacy policy
Google Analytics 4
Provider: Google Ireland Limited, Gordon House, Barrow Street, Dublin 4, Ireland
Purpose: Usage statistics: page views, clicks, scroll depth, sources, devices, load times; in the shop also basket, checkout, purchases, registration and sign-in
Cookies and storage: _ga, _ga_<ID> (2 years)
Privacy policy

Marketing

Measures whether our ads on Facebook, Instagram and TikTok lead to visits and downloads on webdonkey.io and to registrations and purchases in the shop, and builds audiences for ads. Providers: Meta and TikTok, with whom we are jointly responsible for this. With this consent Google may also set advertising cookies (_gcl_…). Transfer to the USA and other countries outside the EU possible. Your choice applies to webdonkey.io and the shop together.

Google Tag Manager, Meta Pixel (Facebook, Instagram), TikTok Pixel
Google Tag Manager
Provider: Google Ireland Limited, Gordon House, Barrow Street, Dublin 4, Ireland
Purpose: Loads the services you agreed to, here and in the shop. Sets no cookies of its own.
Cookies and storage: none
Privacy policy
Meta Pixel (Facebook, Instagram)
Provider: Meta Platforms Ireland Limited, Merrion Road, Dublin 4, D04 X2K5, Ireland
Purpose: Ad measurement and target groups on Facebook and Instagram, also in the shop (basket, checkout, purchases, registration)
Cookies and storage: _fbp (3 months), _fbc (3 months, only if you arrive through an ad), fr on facebook.com (3 months)
Privacy policy
TikTok Pixel
Provider: TikTok Technology Limited, 10 Earlsfort Terrace, Dublin, D02 T380, Ireland
Purpose: Ad measurement and target groups on TikTok, also in the shop (basket, checkout, purchases, registration)
Cookies and storage: _ttp, _tt_enable_cookie (13 months)
Privacy policy