Give your LLM visibility.
Allow them to see what you see.
A macOS menu bar app and an MCP server. You ask your assistant about something on your screen. It takes one screenshot through this tool and answers with your screens in view.
What it does.
Any MCP-capable assistant
It is an MCP server. If your client speaks MCP, it can use this. No per-client plugin, no browser extension.
One capture, on request
The model asks for a screenshot and gets one. It does not watch continuously, does not draw on your screen, and does not automate anything.
Serving is a switch
Start and Stop live in the menu bar. With serving off, every session fails fast with stopped, whatever the client asks for.
Permission state is visible
macOS requires the Screen Recording grant. The menu bar shows whether you have given it, requests it when you have not, and opens the right System Settings pane when macOS will no longer ask.
Setup written for the model
Copy setup instructions puts a block on your clipboard addressed to your assistant. Paste it in and the assistant edits its own MCP config. Your client knows its own config format better than we do.
Test connection
Runs one capture through the same path a client uses and reports what happened. Prove it works before you blame your LLM.
How it works.
- 1. Install the app. It sits in the menu bar and owns the Screen Recording grant.
- 2. Click Copy setup instructions and paste the block into your assistant. It edits its own MCP config.
- 3. Your client spawns a small stdio shim per session. The shim talks to the app over a local socket. The app is the only process macOS grants screen access to.
- 4. Ask your assistant something about what is on screen. It requests a capture and answers with it in view.
It captures your screens. That is the whole product.
Worth being direct about, because it is the part you should think about before turning it on.
macOS gates it. Nothing is captured without the Screen Recording grant. You give that grant explicitly and you can revoke it in System Settings at any time.
You gate it too. Stop serving from the menu bar and no capture happens, no matter what any connected client asks for. Every session fails fast with stopped.
The screenshot goes to the assistant that asked for it. If that assistant runs in the cloud, the contents of your screen go to the cloud. That is inherent to handing an image to a model, not something this app can undo. Point it at a local model and the image stays local.
Senses.
DesktopAudio gives an assistant hearing. DesktopVisionMCP gives it sight. Same family, same idea: give the model one more channel onto what is already happening on your machine, and nothing else.
Pricing.
- One-time purchase. No subscription.
- No trial.
Download.
Setup lives inside the app. Once it is running, Copy setup instructions and Test connection do the rest, so there is nothing to read here first.
Download coming soonQuestions.
Which assistants does it work with?
Any client that speaks MCP. The server is the interface; nothing here is specific to one vendor.
Does it draw on my screen?
No. It captures. It does not overlay, annotate, or render anything on your display.
Does it watch my screen continuously?
No. One screenshot per request, and only while serving is on.
Can it click or type for me?
No. It does not act on your machine. It answers one question: what is on screen right now.
Is there a trial?
No. It is a one-time purchase.
Who is behind it?
Karma Digital, the same people who make DesktopAudio.