AI Assistant
Markpion can ask Claude, Anthropic's AI model, to work on your text: improve it, fix its spelling and grammar, shorten it, summarize it, translate it, continue it, or follow your own instruction. It's off by default, uses your own Anthropic API key, and never changes your document until you've reviewed the answer.
The assistant is part of the desktop app. The browser demo on this website doesn't include it.
Set it up
- Create an API key in the Anthropic Console (API keys → Create key). Anthropic bills your account for what you use.
- In Markpion, open Settings → AI Assistant and turn on AI commands.
- Paste the key and choose Save Key. Markpion checks it with Anthropic, then stores it in Windows Credential Manager or the macOS Keychain (on Linux, in a file only your account can read, in Markpion's settings folder). It's never written to Markpion's settings file.
- Optionally choose the model: Claude Opus 5.5 (most capable, the default), Claude Sonnet 5.5 (faster and cheaper), or Claude Haiku 4.5 (fastest and cheapest).
To stop using it, turn AI commands off, or choose Remove Key.
Use a local model instead
Instead of Claude, the commands can use a model running on your own computer through Ollama. Your text then stays on your computer, and no API key or account is needed.
- Install Ollama and download a model, for example
ollama pull llama3.2in a terminal. Ollama runs in the background athttp://localhost:11434. - In Settings → AI Assistant, turn on AI commands and choose A local model with Ollama.
- Markpion lists the models Ollama has; pick one. If Ollama runs on another port, change the Ollama address and choose Connect. Only addresses on this computer (
localhost,127.0.0.1,[::1]) are accepted.
The same commands, review and Stop work as with Claude. Local models are usually slower (the first request also loads the model) and their answers less reliable, depending on the model and your computer.
Commands
The AI menu, and the command palette (search for "AI"), have these commands. They work on the selected text; with nothing selected, on the paragraph at the cursor.
| Command | What Claude does | The answer goes |
|---|---|---|
| Improve Writing | Makes the text clearer, more concise and more natural, and fixes mistakes | Over the original |
| Fix Spelling and Grammar | Fixes spelling, grammar and punctuation only | Over the original |
| Make Shorter | About half as long, keeping the key points | Over the original |
| Summarize | A few sentences, or a short list | Below the original |
| Continue Writing | Writes one or two more paragraphs from the cursor, using up to 6,000 characters before it | At the cursor |
| Translate… | Translates into the language you type | Over the original |
| Ask Claude… (Ctrl+J / Cmd+J) | Follows your own instruction, such as "turn this into a table" | Over the original |
| Write… (Ctrl+Shift+J / Cmd+Shift+J) | Writes new text from your instruction alone, such as "an introduction about our release process". No document text is sent | At the cursor |
The review window opens straight away and shows the answer as Claude writes it. Stop (or closing the window) ends the request, so Claude stops writing and nothing is applied.
Review before anything changes
For commands that replace the text, the window shows the Changes first: the original and the suggestion as a diff, with the changed words marked, so a grammar fix is easy to check. It follows your edits to the suggestion. Choose Original to see the original text as it was.
When Claude has finished, you can edit the suggestion, then:
- Replace (or Insert / Insert Below, depending on the command) puts it into the document as one edit, so Ctrl+Z undoes it;
- Copy copies it to the clipboard;
- Discard closes the window without changes.
If you changed the original text while Claude was working, Markpion doesn't overwrite it; copy the suggestion or run the command again.
AI can make mistakes: always check the suggestion before you use it.
Privacy
- With a local model, nothing leaves your computer: the text goes to Ollama at the address you set, which must be on this computer, and there is no consent prompt. The rest of this section is about Claude.
- Nothing is sent until you run an AI command. The first time, Markpion asks you to confirm. After that, running a command is your consent to send that text.
- What's sent: the selected text or paragraph (or, for Continue Writing, the text before the cursor; for Write, nothing but your instruction) and the command's instruction, directly from your computer to Anthropic's API (
api.anthropic.com), under your key. Nothing else from your documents or your computer is sent. - How Anthropic handles it is set by your agreement with Anthropic and its privacy policy.
- Markpion's diagnostic log records only that a request was made, its size and whether it worked, never the text or the key.
- The key is only read by Markpion's native part; the editor interface never sees it.
Troubleshooting
| Message | What to do |
|---|---|
| "The API key was rejected" | Check the key in the Anthropic Console, then save it again in Settings |
| "Your Anthropic account can't be billed" | Add credit or a payment method in the Anthropic Console |
| "Too many requests" or "Claude is busy" | Wait a moment and try again (Markpion retries twice automatically) |
| "The answer was too long" or "The selected text is too long" | Select less text (a request can include up to 200,000 characters) |
| "Couldn't reach the Claude API" | Check your internet connection and any proxy |
| "Claude declined this request" | Reword the instruction, or select different text |