Windows 11 Clipboard AI: Sirius Tray Assistant Connects to Self-Hosted Models

The appeal is refreshingly specific: not an AI assistant promising to reorganize your entire working life, but a tray utility that helps with small, repetitive jobs. Pande’s experience also highlights two details worth getting right before trying it: where the model actually runs, and what happens to the material you send it.

A clipboard assistant, not a replacement clipboard​

Pande’s setup pairs Sirius, developed by zoott28354, with Gemma 4 E4B in Q4_K_M format, served through llama.cpp. He describes invoking clipboard analysis, adjusting prompts for individual tasks, and returning to saved conversations for follow-up.

Sirius’s repository independently documents clipboard-text analysis, native Windows screen capture, customizable tray prompts, and persistent conversations. It supports multiple inference backends, including llama.cpp, Ollama, and LM Studio. The documentation inspected identifies version 3.6.0; that does not establish which version Pande used.

The distinction matters: this is an application consuming copied material and passing it to a configured model, not a new intelligence layer built into Windows itself.

The useful part is the prompt​

Pande reports using custom instructions to strip formatting from copied PowerShell scripts and Linux commands, summarize code, and translate foreign-language material into English. He also uses screenshot analysis when text cannot be selected directly.

Those are sensible targets for a narrowly focused assistant. However, a productivity prompt should define what must not change as carefully as what should.

For example, these are suggested instructions—not tested results:

• Translation: “Translate into English. Preserve commands, file paths, product names, and version numbers exactly.”
• Command explanation: “Explain what this command does, identify destructive operations, and do not execute or rewrite it.”
• Screenshot transcription: “Transcribe the visible text. Mark uncertain characters rather than guessing.”

For code cleanup, request exact preservation of characters and line breaks. Treat the output as a proposed transformation, not an authoritative replacement for the original.

The server address is an important setup detail​

Pande connected Sirius to [url]http://localhost:8080/v1[/url]. That address belongs to his configuration, not a universal Sirius requirement: the repository lists port 8033 for its llama.cpp preset and 8080 for Llama-Swap, with editable endpoints. Match the address to the running server.

He also describes using a Custom profile to reach llama-server inside a Proxmox LXC, freeing his daily-driver GPU. Sirius documents LAN connections, provided the backend listens on a reachable network interface and its firewall permits access. Here, localhost means the Windows computer running Sirius—not another machine.

The practical takeaway: establish a working backend first, then connect the clipboard utility to it. Installing a tray application is not the same as installing and serving a model.

Small-model performance needs some qualification​

Pande reports more than 110 tokens per second on an RTX 3080 Ti, plus acceptable everyday performance on a GTX 1060 laptop and a GTX 1080-backed container. These are personal observations, not standardized benchmarks: his account does not specify enough measurement conditions to predict another reader’s results.

Google describes E4B as an effective 4B-parameter model, rather than treating the name as a direct measure of its complete memory footprint. Google also confirms image-input support across the Gemma 4 family, supporting the model choice for screenshot analysis.

Memory planning deserves equal attention. Google estimates approximately 4.5 GB for E4B at Q4_0, but that is not an exact requirement for Pande’s Q4_K_M build. Its estimates exclude additional memory for supporting software and the context cache; requirements vary with the inference environment and conversation length.

In short, “small enough for local use” does not mean “guaranteed to fit comfortably on every older GPU.”

Local inference still leaves a data trail​

Sirius documents local SQLite conversation storage under %AppData%\Sirius AI Tray Assistant. It also sends conversation history to the selected backend. Consequently, persistent chat is both a convenience and a retention consideration; selecting another inference host changes where processing occurs.

Before analyzing confidential material, verify the destination and decide whether that content belongs in saved history. “Self-hosted” should describe a known data path, not serve as a substitute for checking one.

A modest use case with a credible payoff​

The strongest lesson from Pande’s experiment is not that a local model can replace every cloud service. It is that clipboard-sized tasks offer a practical place to start: short inputs, explicit instructions, and results a person can inspect immediately.

Keep that inspection step. Translation, screenshot transcription, and command explanations are useful assistance—not permission to paste blindly into an elevated terminal. The clipboard may have gained a brain, but it still benefits from adult supervision.