PowerToys Advanced Paste AI: Local or Cloud Models?

PowerToys Advanced Paste is the closest thing Windows has to a first-party AI clipboard feature. It can take the current clipboard content and rewrite it as Markdown, JSON, plain text, code, or a custom prompt, and it can do this with either a local model or a cloud model. The choice matters: the local path keeps the clipboard on the device; the cloud path sends it to a vendor. This guide walks through both setups, the trade-offs, and when each is appropriate. For related reading, see AI clipboard managers: useful or a new leak, Windows Recall, Click to Do, and copied text, what local-first software means for clipboard apps, and why some clipboard apps will never sync.

What Advanced Paste actually does

Advanced Paste is a PowerToys module that adds paste-time transformations to the Windows clipboard. The default hotkey is Win + Shift + V, which opens a small menu of transformations:

• Paste as plain text — strips formatting.
• Paste as Markdown — converts rich text or HTML to Markdown.
• Paste as JSON — extracts structured data from text.
• Paste as code — wraps the clipboard in a code block, with optional language detection.
• Custom paste — a free-form prompt: "translate to French," "summarise in two sentences," "extract the URLs."

The first three are local operations that do not call any model. The code and custom paste features can use either a local or a cloud model. This is where the privacy decision sits.

The two model paths

Local model

Advanced Paste can call a local model through Ollama or Foundry Local. The clipboard content stays on the device; the model runs locally; no data is sent to a vendor. The setup:

1. Install Ollama or Foundry Local. Both are free. Ollama is the more common choice; Foundry Local is Microsoft's own local model runtime.
2. Pull a model. For Ollama, ollama pull llama3.1:8b or ollama pull qwen2.5:7b are reasonable starting points. RAM use is 4–8 GB for a 7B model.
3. Configure Advanced Paste. In PowerToys Settings → Advanced Paste, point the local model endpoint at the Ollama or Foundry Local address (typically http://localhost:11434 for Ollama).
4. Test. Copy a paragraph, press Win + Shift + V, choose "Custom paste," type "summarise in one sentence," and confirm the response comes back.

The local path's strengths are privacy and offline operation. Its weaknesses are speed (a 7B model on a laptop takes a few seconds per request), quality (local models are good but not GPT-4-class), and RAM cost (the model resident in memory eats 4–8 GB that other apps cannot use).

Cloud model

Advanced Paste can also call the OpenAI API, with an API key the user provides. The clipboard content is sent to OpenAI, processed, and the response is returned. The setup:

1. Get an OpenAI API key. This requires an OpenAI account and a payment method.
2. Configure Advanced Paste. In PowerToys Settings → Advanced Paste, paste the API key.
3. Test. Same as above, but the response comes from OpenAI.

The cloud path's strengths are speed (sub-second for small inputs), quality (GPT-4-class models are better than local 7B models), and RAM cost (zero, beyond the PowerToys overhead). Its weakness is privacy: every Advanced Paste call sends the clipboard content to OpenAI, which retains data per its policy unless the user opts out of training.

When each is appropriate

The honest split:

• Local for anything that might contain secrets. API keys, connection strings, internal URLs, code from proprietary repos, customer data. If you would not paste it into ChatGPT manually, do not let Advanced Paste send it to OpenAI.
• Cloud for content you would share publicly. A blog post draft, a public doc, a tweet you are about to publish. The cloud model is faster and better, and the privacy cost is bounded to content you were going to share anyway.
• Neither for content that is genuinely sensitive but public-shaped. A draft press release, an internal memo that reads like a public doc. The safe default is local, even when it is slower.

For users who handle production credentials, code from proprietary repos, or customer data, the answer is local-only. The cloud path is a leak waiting to happen, because the muscle memory of "press Win+Shift+V, choose custom paste" does not distinguish between sensitive and non-sensitive content. A single slip sends an API key to OpenAI.

A summary table

The last row is the underrated option. The non-AI transformations in Advanced Paste — paste as plain text, paste as Markdown, paste as JSON — are local, deterministic, and instant. For users whose main use case is stripping formatting or converting rich text to Markdown, no model is needed at all.

Configuring for safety

For a privacy-conscious setup:

1. Leave the OpenAI API key field empty. This disables the cloud path entirely.
2. Install Ollama and pull a 7B model. Configure Advanced Paste to use it.
3. Set the custom paste prompt to a sensible default. "Summarise the clipboard in one sentence" is a useful default; "rewrite as a tweet" is another.
4. Test before relying on it. Copy a paragraph, run the transformation, confirm the response is reasonable.

This setup gives you the AI features without the leak. The cost is speed: a 7B model on a typical laptop takes 2–5 seconds per request, which is slower than the cloud path but tolerable for occasional use.

For more on the broader AI-on-the-clipboard question, see AI clipboard managers: useful or a new leak and should clipboard apps cluster items with embeddings.

The cold-start cost

The local model has a one-time cold-start cost the cloud path does not. The first Advanced Paste call after Ollama has been idle for a while triggers the model to be loaded back into VRAM or system RAM, which can take 5–15 seconds on a laptop. Subsequent calls are fast. The workaround is to keep Ollama's keep-alive high enough that the model stays resident during the workday.

How this relates to the broader clipboard stack

Advanced Paste is not a clipboard manager. It is a paste-time transformer that operates on the current clipboard content. To get history, search, and pinning, you still need Win+V, Ditto, CopyQ, or Edge-Drop. The realistic stack for a Windows user who wants AI features:

• Win+V or Ditto for history.
• Edge-Drop for spatial access and drag-out.
• PowerToys Advanced Paste with a local model for AI transformations.

Three tools sounds like a lot, but each does one job well, and the total RAM is reasonable (Ditto is light, Edge-Drop is 130–160 MB, Ollama with a 7B model is 4–8 GB if and only if you are using it). For users who do not want AI features, drop the third tool and the RAM cost goes with it.

For more on stacking clipboard tools, see can you run two clipboard tools at once and PowerToys as a clipboard strategy.

What to watch for

The category is moving fast. Things worth tracking:

• Windows Copilot Runtime. Microsoft's OS-level AI surface, which may eventually let Advanced Paste call a bundled local model without requiring Ollama. As of mid-2026, this is not the default; check the current PowerToys release notes.
• Local model quality. A 13B local model is meaningfully better than a 7B, at the cost of more RAM. Watch for laptop-class hardware that can run a 13B comfortably.
• OpenAI data retention. OpenAI's policy on training data may change. If you use the cloud path, re-check the policy periodically.
• PowerToys release cadence. PowerToys ships frequently, and the Advanced Paste settings UI has changed between releases. Verify the steps above against the current version.

For most users, the practical takeaway is: install Ollama, configure Advanced Paste to use it, leave the OpenAI field empty, and accept the speed cost. The privacy win is worth the wait.

Related reading

• Windows Recall, Click to Do, and Copied Text
• Should Clipboard Apps Cluster Items With Embeddings?
• What Local-First Software Means for Clipboard Apps
• How to Enable Clipboard History in Windows 11

Sources

• Microsoft Learn — PowerToys Advanced Paste — official documentation for the Advanced Paste module, including local and cloud model configuration
• Microsoft Learn — Foundry Local — official documentation for Microsoft's local model runtime, an alternative to Ollama
• Ollama — GitHub README — official documentation for the local model runtime most often used with Advanced Paste
• OpenAI — API data usage policies — official OpenAI policy on how API-submitted data may be used, relevant when deciding whether to use the cloud path
• Microsoft Learn — Windows Copilot Runtime — official documentation for the OS-level AI surface that may eventually provide a bundled local model

Copy. Stack. Drop.

Transform your clipboard into an interactive edge shelf. Stack, pin, and drag assets into any app with zero friction.

Download for Windows Get from Microsoft Store

How to Install Guide · First 10 Minutes Guide · Drag & Drop Guide · Edge-Drop vs Win+V · Support