daniel-farina/hipdash

Hipdash is a live monitoring dashboard for MTPLX, which the README describes as a multi-token-prediction LLM runtime, a system for running large language models locally with a technique for predicting multiple output tokens at once to speed up text generation. The dashboard streams real-time metrics from a running MTPLX instance across several views. The Overview tab shows key figures: tokens decoded per second, time to first token (TTFT), CPU and memory usage, and session bank state. The MTPLX tab goes deeper with a live visualization of the generation pipeline and 12 history charts tracking metrics like decode speed, prefill time, accepted tokens, cached tokens, and context size over time. A Computer tab shows host-level stats (CPU, memory, swap, disk, and thermal readings), and a Restarts tab logs every restart event with selectable history. An OpenCode tab displays activity from OpenCode agents, AI coding assistants, pulling their configuration and recent tool usage from a local config file at ~/.config/opencode/opencode.json. Everything is persisted to a local SQLite database, so the dashboard keeps history across restarts and can show benchmark run boundaries. Old data is automatically pruned on a 14-day retention policy. The backend polls MTPLX every 1.5 seconds and a sidecar service every 3 seconds, writing snapshots and time-series data to SQLite. A live SSE stream (Server-Sent Events, a way for a server to push updates to a browser in real time) taps completions traffic. The backend is built with Node.js, Express, and better-sqlite3. The frontend uses Vite, React 18, and TypeScript.