User Interface
Settings

Inspiration

When something goes wrong with a computer, the fix usually means searching forums and pasting commands you don't understand. For non-technical people, that's stressful and risky. We wanted an assistant that does the technical digging for you and answers in plain language — and, through Alexa+, could one day just tell you how your computer is doing out loud.

What it does

SysPilot is a system health assistant for everyday users. You ask a question in normal words, and it inspects your machine — CPU, memory, battery, drivers, startup apps, network, temperatures, disk health, and more — then answers like a helpful friend, not a manual.

Diagnoses in plain language. "Why is my computer slow?" returns a simple, jargon-free explanation of what's actually using your resources.
Knows what's safe. Ask which apps you can close, and it separates harmless background apps from system services you shouldn't touch.
Fixes only with your approval. When a fix would change your system, SysPilot never runs it silently. It shows the exact command and waits for an Allow / Not now button. Anything needing admin rights is handed to you as a copy-and-run command, so you stay in control.
Remembers your conversations. A ChatGPT-style interface saves every chat, so SysPilot becomes a running record of your computer's health.

How we built it

SysPilot is built around a self-hosted Model Context Protocol (MCP) server exposed over Streamable HTTP. The server publishes 39 read-only diagnostic tools plus approval-gated action tools. An agent loop runs a tool-capable LLM (Groq) against these tools, so every factual claim comes from a real tool call — the model never invents system details.

Backend: Python, FastAPI, the MCP Python SDK, served with uvicorn.
Diagnostics: standard Linux interfaces and utilities, wrapped into safe, structured tools.
Frontend: a React + TypeScript desktop app (pywebview) with a ChatGPT-style UI and local, persistent chat history.
Safety: state-changing actions are staged and require explicit on-screen approval; privileged commands are never auto-run.
Alexa+: the same MCP tool surface is what Alexa+ connects to — Alexa+ handles the speech, SysPilot does the diagnosing.

Challenges we ran into

Making highly technical system data genuinely understandable to a non-technical person, without dumbing it down to the point of being useless.
Keeping the agent honest: forcing it to call a tool before stating any fact, instead of guessing.
Designing a safety model that never runs risky commands on its own while still feeling effortless — a button-based approval gate and copy-and-run cards for anything needing admin rights.
Working within the hackathon's Alexa+ constraints: the official Alexa+ tooling is in preview and unavailable to participants, so we built against the sanctioned simulated path using a standard MCP server.

Accomplishments that we're proud of

39 working diagnostic tools covering the whole machine, from CPU and memory to battery, GPU, drivers, and disk health.
A genuinely non-technical experience that explains problems in plain words.
A safety-first design where nothing changes without the user's approval.
A clean, familiar ChatGPT-style interface with persistent conversations.

What we learned

Plain-language translation is a design problem as much as a technical one.
Grounding an agent in real tool calls is the difference between a trustworthy assistant and a confident guesser.
MCP is a clean, portable way to expose the same capabilities to both a desktop app and a voice assistant like Alexa+.