AI Actions Guide: Set Up a Model, Use It, Make It Yours
A step-by-step guide to AI Actions on selected text: connecting a model, running your first action, writing your own, and fixing the things that go wrong
A step-by-step guide to AI Actions on selected text: connecting a model, running your first action, writing your own, and fixing the things that go wrong. For the short pitch, see the feature page.
AI Actions is off until you set it up, and nothing is sent anywhere before that.
1. Find the Setup Window
Tray menu → AI and Meetings. The menu holds:

| Entry | What it opens |
|---|---|
Ask AI About the Selection (Ctrl+Alt+Shift+A) | The command palette in AI mode, with your selection captured. |
| My AI Actions... | The editor for actions of your own. |
| AI Per-Program Defaults... | A usual action (and optionally a model) for each program. |
| AI Actions Setup... | Where you connect a model. This guide's main window. |
| Model | A submenu that switches the model in one click (the one in use is ticked). |
| Meeting Notes (offline transcript) | (SlyPilot Pro) See the Meeting Notes guide. |
The Welcome window has an AI actions... button that opens the same Setup window. You can also run SlyPilot.exe ai-setup from a command line.
2. Choose One of Three Ways to Get a Model
| If you... | Choose | Your text goes to |
|---|---|---|
| ...have nothing installed and want it simple | Local model... (section 3) | Stays on your PC |
| ...already run Ollama or LM Studio | Ollama (PC) or LM Studio (PC) quick-fill | Stays on your PC |
| ...have an Ollama cloud or OpenRouter key | Ollama cloud or OpenRouter quick-fill | The service you named, after your tick |
3. Option A: The Built-In Model (No Server, No Account)
- In AI Actions Setup... press Local model.... The window Built-in model on this PC opens.
- Choose the model in the Model list: Qwen3.5 0.8B (508 MB) for most PCs. A stronger PC (16 GB of memory and 12 or more threads) is also offered the 2B model (1.2 GB, slower, a little better). A PC with under 6 GB of memory is offered none.
- Press Download and use. A consent window lists every file to download, its size and where it comes from (the llama.cpp program, 19 MB, from github.com; the model from huggingface.co). Nothing is downloaded until you click Yes, and each file is checked against its published SHA-256 checksum before it is used.
- When it finishes, the window says it is in use. Close it and try an action (section 6).
What to expect:
- It starts when you ask for something (a few seconds the first time; the card says "Starting the local model"), then answers. It uses about 0.9 GB of memory while running.
- Free the memory after: 5 minutes, 10 minutes (the default), 30 minutes or Never.
- It is good at fixing grammar and spelling and cleaning dictated text; shortening and rewriting in another tone are weaker than with a large model. English only. Selections over about 3,000 characters are refused with a sentence.
- It runs on your processor, so answers take seconds. Ollama or LM Studio with a graphics card will be much faster.
- Stop using it switches back to a server; Remove from this PC deletes the downloaded files (
%LOCALAPPDATA%\SlyPilot\LLM) and you can download again later. - The same model also handles dictation polish and meeting summaries (both are SlyPilot Pro features; AI actions themselves are free).
4. Option B: Your Own Server, or a Cloud Service
In AI Actions Setup:
- Quick fill (press one, or type any OpenAI-compatible address in Address):
| Button | Address | Notes |
|---|---|---|
| Ollama (PC) | http://127.0.0.1:11434/v1 | Install Ollama, run ollama pull <model> once, leave Ollama running. |
| LM Studio (PC) | http://127.0.0.1:1234/v1 | Load a model and start LM Studio's local server (Developer tab). |
| Ollama cloud | https://ollama.com/v1 | Paste your Ollama API key into API key. The model list works even before you paste it. |
| OpenRouter | https://openrouter.ai/api/v1 | One key from openrouter.ai/keys (it starts with sk-or-) reaches many models, many of them free (names ending :free). |
- Press Load models. The models the server offers appear in the On server list; pick one and its name goes into Model (you can also just type a name). For short edits a small, fast model is best; reasoning models work too.
- For an address outside your PC (Ollama cloud, OpenRouter, any
https://service): tick the box Allow sending my selected text and instructions to . Without the tick nothing is sent. The card names where each request went. - Press Test connection. It asks the model a tiny question, so it checks both the model name and the key (the Ollama cloud model list works without a key, so the list alone proves nothing).
- Tick Turn AI actions on and press Save. The key box is emptied after Save on purpose; the line under it says a key is stored. Remove key deletes it.
Answers: the drop-down offers Show the answer on a card first (recommended) or Replace the selection right away for fix, polish, shorten, translate. Even with the second choice, an answer that looks wrong is shown on the card instead of replacing your text.
OpenRouter specifics
- Load models lists free models first (the router
openrouter/freefirst, then the families that do plain text edits well, then the rest) and leaves out models that make no text. - Free models have strict limits: a few requests per minute and a daily cap (more once your account holds some credit).
- Tick OpenRouter: if my free model is busy, let OpenRouter try up to two other free models to let the service move down a list when yours is rate-limited or down (the card says which model answered). Untick it if you want only the model you picked.
- Each model may belong to a different company, and the text goes to whichever one answers. Free models may keep or learn from prompts depending on your OpenRouter privacy settings: read them first.
- SlyPilot switches the thinking of OpenRouter models off by default (
[SlyAction] OpenRouterReasoning=off;low,excludeanddefaultare the other choices), because thinking costs time and some free models then put their analysis into the answer.
5. Voice and Palette Cheat Sheet
| You say or type | What happens |
|---|---|
Ctrl+Alt+Shift+A, then fix | Fix spelling and grammar in the selection. |
a polish / a shorten / a formal / a casual | Clearer; shorter; more formal; more casual. |
a translate german | Translate to a language you name. |
a summarize / a explain / a draft reply / a list facts | Answers on the card. |
a <anything in your own words> | A free instruction on the selected text. |
... @screen / @clip / @meeting | Read from a dragged screen region (OCR), the clipboard, or your newest meeting notes. |
... >replace / >insert / >copy | Use the answer directly instead of the card. |
f <question> | Follow up on the last answer (only while the thread is alive). |
| "fix that", "polish that", "make that shorter", "make that formal", "make that casual", "summarize that", "explain that" | The same actions by voice. |
| "ask AI", "ask AI about this screen", "ask AI about my meeting" | Listens for one request. |
| "use the answer", "copy the answer", "discard the answer" | Works the card. |
| "follow up" | Then say your follow-up question. |
| "put that back" | Undoes the last replace or insert (checked first). |
The voice phrases with "that" need a clearly recognised voice. A bare a in the palette lists your usual actions for the program in front. The full list of voice commands is in the Voice Command Reference.
6. Your First Action, End to End
- Type a sentence with a typo in any text field, and select it.
- Press Ctrl+Alt+Shift+A. The palette opens with
atyped. - Type
fixand press Enter. - The card appears. Read the answer, then press Replace selection.
- Changed your mind? Say "put that back" (or type
a undo).
If nothing happens, jump to Troubleshooting.
7. Write Your Own Actions
Tray → AI and Meetings → My AI Actions... (or the My actions... button in AI Actions Setup, or SlyPilot.exe ai-actions).

Each action has:
- Name(s): several names can be separated by commas,
- Prompt: tells the model what to do with the selected text,
- Answer: Show on a card, Replace the selection, Insert after the selection or Copy to the clipboard,
- optionally the tick Also works when nothing is selected.
The buttons are Add as new, Update selected, Remove and Close. A Per program... button opens the defaults window of the next section.
Examples:
| Name | Prompt | Answer | Option |
|---|---|---|---|
reply, respond | Write a short, polite reply to this message. | Card | |
bullets | Turn the text into short bullet points. | Replace | |
ideas | Give three title ideas for this text. | Card | Also works when nothing is selected |
Then a reply (or a bullets, a ideas) runs it, just like a fix.
Rules: a name cannot be one of the built-in action words (fix, polish, shorten, formal, casual, translate, summarize, explain) and must not start like translate . A row that breaks a rule is refused with the reason and nothing is saved. The selected text always goes to the model as data between tags, never as instructions.
Where they live: in the settings file under [SlyActions], one line each, names = prompt | delivery | free (\n is a new line, \p a | sign, \\ a backslash), so they are easy to back up or edit by hand.
8. A Usual Action and Model for Each Program
Tray → AI and Meetings → AI Per-Program Defaults... is a list of Program, Usual action and Model, with the buttons Add as new, Update selected and Remove. Every change is saved at once. The Program box offers the programs that have a window open now; the program name is the process name without .exe (see Task Manager). Fill in the Usual action (what you would type after a) and, optionally, a Model (leave it empty for the model chosen in AI Actions Setup).

Or in the settings file under [SlyActionApps]:
outlook = draft reply
chrome = summarize | gpt-oss:120bWith that program in front, a bare a in the palette lists Your usual here: ... first, and requests use that program's model. The built-in model has only one model, so a model entry changes nothing for it. The address, key and outside-service permission never change per program.
A model for each job: [SlyAction] PolishModel is the model for dictation clean-up and [SlyAction] MeetingModel the one for meeting summaries; both default to the model chosen in Setup.
9. Optional Extras
Polish what you dictate (SlyPilot Pro)
This belongs to AI dictation, which is part of SlyPilot Pro; AI actions themselves stay free. In AI Actions Setup, tick Polish dictated text with this model. Off by default. Your raw text is always the fallback: if the model is slow, fails or answers something that does not look like your words, the raw text is typed and the HUD says why; Esc while it says "Polishing..." types the raw text at once. Dictations of fewer than 8 words are typed as they are. A model outside your PC needs its own second tick (Also polish dictated text with ). More in Dictation Guide.
Follow-ups
With an answer on the card, press Follow up... (or say "follow up", or type f and your question). Up to 6 follow-ups, kept in memory for 10 minutes, gone after Discard.
Screen text and meeting notes
@screen asks you to drag a rectangle around any text on screen. For a model outside your PC it asks once each time SlyPilot starts; the tick Do not ask before sending text read from the screen to stops that question for that host. @meeting (it reads your newest meeting notes, a SlyPilot Pro feature) asks every time for a model outside your PC.
A Touch Bar button
Settings → Touch Bar → custom buttons: pick an AI preset (fix, polish, explain, ask about text on the screen), or choose the type ai and write what you would type after a (for example shorten or summarize @clip).
From your phone (SlyPilot Pro)
With the Phone Remote (SlyPilot Pro, Chrome on Android), its AI tab runs the same actions on the PC's selection. See the Phone Remote guide.
What it costs
AI Actions Setup shows how many requests went to outside services today and in the last 30 days (with a rough token count) and how many were answered on your PC. Only counts are kept, never text, for 62 days.
Settings-file options
Settings live in SlyPilot.ini under [SlyAction]. The ones people reach for:
| Key | Default | What it does |
|---|---|---|
MaxChars | 12000 | The longest selection that is sent (3,000 for the built-in model). |
TimeoutSec | 60 | How long to wait for an answer. Raise it for a slow model. |
Streaming | smart (1) | always streams every card answer, 0 never. |
FallbackLocal | 1 | 0 turns off the retry on the built-in model. |
LocalIdleMin | 10 | Minutes before the built-in model stops (0 = never). |
See the Configuration Guide for how the settings file works. AllowRemote and the encrypted key are written by the Setup window only.
10. Troubleshooting
| What you see | What it means and what to do |
|---|---|
| "Not set up yet: opening AI Actions Setup" | No model is chosen or AI actions are not turned on. Finish steps 4 and 5 of section 4, or use the built-in model. |
| "Nothing is selected: select some text first." | The action needs a selection. Select text first, or use @screen / @clip. |
| "Still working on the last request: click Cancel on its card first." | The model is busy with the previous request. Press Cancel on its card (or the hotkey while it is thinking), then ask again. |
| "Could not reach the server. Is Ollama / LM Studio running, and is the address right?" | Start the program, check the address in Setup, or use Test connection. If the built-in model is downloaded, SlyPilot may answer on your PC instead. |
| "The model did not answer in time." | A slow computer or a large model. Raise TimeoutSec, or choose a smaller model. |
| "The server did not accept the API key (401 / 403)." | Paste the key again in Setup and press Test connection. |
| "The server answered 404: the address or the model name is not right." | Use Load models and pick a name from the list. |
| "...needs credit (402)." | With OpenRouter pick a model whose name ends with :free, or add credit to your account. |
| "...too many requests (429)." | Free models have strict limits. Wait a moment, pick another model, or press Try another model on the card. |
| "This model wrote out its thinking instead of the answer." | Some free models do. Press Try another model, or choose another one in Setup. |
| "The model only produced its reasoning and no answer." | Pick a model without a thinking mode, or allow more time. |
| "The model returned an empty answer." | Try again or switch model. |
| "That is N characters; the limit is..." | Select less, or raise MaxChars. |
| "This address is outside your PC. Open AI Actions Setup and tick the box that allows it..." | Tick the allow box for that host, or use a model on your PC. |
| "Not sent: text read from the screen only goes to after you agree." | Answer Yes to the question, or use a model on your PC. |
| Nothing is copied from a terminal | Terminals have their own copy key, so SlyPilot does not copy a selection from them. A free question is still answered on the card. |
| Nothing is read from a browser window | Private browsing and password-manager windows are never read or written. |
| The answer looks wrong | Small models make mistakes. Check the card before you replace, or use a stronger model. Nothing changes until you click. |
Still stuck? Tray → Help and Backup → Check my setup also names a failed last AI request (a wrong key, a server that is not running) and has a button that opens the setup. See Troubleshooting & the Doctor.
11. Privacy Checklist
- [ ] Where does my text go? Look at the address in Setup:
127.0.0.1(or the built-in model) means it stays on your PC. Anything else is an outside service. - [ ] Did I tick the allow box on purpose? It names the host. Untick it to stop sending.
- [ ] Do I use a free OpenRouter model for private text? Read your OpenRouter privacy settings first. Prefer a model on your PC for anything private.
- [ ] Do I use a local Ollama that is signed in with a
:cloudmodel? That text also leaves your PC, and the card cannot tell you. - [ ] Dictation polish (SlyPilot Pro) to an outside service? It needs its own tick; leave it off for sensitive dictation.
- [ ] Is my key safe? It is stored encrypted for your Windows user (DPAPI) and is never logged, exported or part of a diagnostics report. Remove key deletes it.
- [ ] Is anything logged? The selected text, instruction and answer are not written to the log.
- [ ] Could the answer run something? No. The answer is only ever pasted as text.
- [ ] Protected windows? Password managers and private browsing windows are never read or written.
Privacy policy → · AI Actions feature page →