So Nvidia app has an opt out MCP server running on your computer.....
retained text
I enabled the MCP server feature in the NVIDIA App version 11.0.9.565 (the setting that lets an AI assistant control the app) and poked at what it actually exposes on my own machine. Sharing what I found to start a discussion, because I'm not sure how comfortable people are with the trade-offs here. What it is Turning on the feature starts NvAppMcpServer.exe and a local HTTP endpoint at [http://127.0.0.1:13508/mcp](http://127.0.0.1:13508/mcp) (served via Windows http.sys). The idea is that an AI client (Claude Desktop, Cursor, etc.) connects and gets tools to read and change NVIDIA App settings. What I confirmed (on my own box, read-only testing) \- A client with no auth token and no Origin header could open a session and list + execute tools. I only ran one read-only call (overlay\_get\_status) — it returned live data: overlay state, mic mode, recording status, screen-capture flags. \- The tools offered to that unauthenticated session include not just reads but actions: launch\_application (starts any app the NVIDIA App has discovered), overlay\_capture (screenshot / toggle recording / save instant replay), plus setting changes (game optimization, battery boost, whisper mode). I did not execute the action tools, so I can't claim they definitely run unauthenticated, but there was no auth gate in front of the read call. \- Read tools leak a decent fingerprint without any auth: exact GPU/VRAM, driver version, your full installed games/apps list, and the currently running game + its PID. What's actually hardened (credit where due) \- Binds to [127.0.0.1](http://127.0.0.1) only not exposed to the network by default. \- Runs as the logged-in user, not admin/SYSTEM. All binaries are validly NVIDIA-signed. \- It does reject browser requests with a bad Origin (so classic DNS-rebinding from a web page is blocked). The gap: a non-browser client that simply sends no Origin passes through. \- launch\_application refuses arbitrary exe paths / shell commands, only known apps. \- There's single-use confirmation-token machinery in the code for mutating calls. The core issue, as I see it "Localhost only" is not a trust boundary against code already running as you. Any local process, a script, an npm package, malware, another app can reach this endpoint without authenticating. So the real framing isn't "someone on the internet," it's: anything running in your user session gets a convenient, unauthenticated API to your GPU stack, your app list, and potentially screen capture / recording. When an AI client is attached, prompt injection is the other path to triggering those actions. Open questions for the thread \- Is an unauthenticated loopback API acceptable for this kind of capability, or should it require a token/pairing by default? \- How do you feel about screenshot/recording being in the exposed tool set specifically? \- Anyone know if NVIDIA documents an auth model (e.g. NVAPP\_MCP\_TOKEN) and whether clients are expected to set it? \- Do you leave it on, or only enable it per-session? (Tested on Windows 11, NVIDIA App, port 13508. Happy to share method. I changed no settings only read.)
document fetched 2026-10-24 10:37 · canonical https://reddit.com/r/nvidia/comments/1xe3mo0/so_nvidia_app_has_an_opt_out_mcp_server_running