Backprop

Run, monitor, and manage ML training runs directly from VS Code.

Backprop is a CLI-first ML trainer with intelligent resource governance — timeboxed runs, GPU monitoring, and auto-resume from checkpoints. This extension brings the full workflow into your editor.

Features

Run Training Scripts

Launch any Python training script directly from the editor. Each run gets its own dedicated terminal — run multiple experiments in parallel while the Governor keeps your machine healthy.

Quick run (Ctrl+Alt+R): One-click launch from any Python file
Configured run (Ctrl+Alt+Shift+R): Set experiment name, timeout, and framework before launching
Play and gear buttons appear in the editor title bar for Python files

Browse all your training experiments in the Backprop activity bar. Each run is expandable to show full details:

Status icons — spinning sync (running), green check (completed), red error (failed), clock (timeboxed)
Expand any run to see: ID, script path, duration, steps completed, current loss, checkpoint count, start/end timestamps
Auto-refreshes when experiments.json changes on disk — no manual refresh needed
Context menus with actions based on run status (stop, resume, show terminal)

Run Management

Stop (Ctrl+Alt+S): Send SIGINT for graceful shutdown — the script can save a final checkpoint before exiting
Resume: Pick up any completed, failed, or timeboxed experiment from its last checkpoint
Show Terminal: Jump to a running experiment's live output

Inline action buttons appear directly on tree items — stop/terminal for running experiments, resume for stopped ones.

GPU Status Bar

A persistent status bar item shows real-time GPU VRAM and temperature, refreshing every 10 seconds. Reads nvidia-smi directly — no CLI dependency required. Click it to launch a configured run.

Keyboard Shortcuts

Shortcut	Command	When
`Ctrl+Alt+R` (`Cmd+Alt+R` on Mac)	Run with Backprop	Python file is active
`Ctrl+Alt+Shift+R` (`Cmd+Alt+Shift+R` on Mac)	Run with Backprop (Configure...)	Python file is active
`Ctrl+Alt+S` (`Cmd+Alt+S` on Mac)	Stop Run	Always

All keybindings can be customized in VS Code's Keyboard Shortcuts editor (Ctrl+K Ctrl+S).

Requirements

Backprop CLI installed globally:
```
npm install -g @mcptoolshop/backprop
```
Python 3.x for training scripts
NVIDIA GPU + drivers for GPU monitoring (optional — extension works without it)

The extension checks for the CLI on activation and offers a one-click install if missing.

Commands

Command	Description	Where
`Run with Backprop`	Quick-launch the active Python file	Editor title bar, Command Palette
`Run with Backprop (Configure...)`	Launch with name/timeout/framework prompts	Editor title bar, Command Palette, Status bar click
`Resume Run`	Resume a stopped experiment	Tree view context menu
`Stop Run`	Send SIGINT to a running experiment	Tree view context menu + inline
`Show Terminal`	Focus a running experiment's terminal	Tree view context menu + inline
`Refresh Experiments`	Manually reload the experiment list	Tree view title bar

How It Works

The extension is a thin UI layer over the Backprop CLI. It reads experiment data from ~/.backprop/experiments.json (the same file the CLI writes to), launches CLI commands in VS Code terminals, and watches the file for changes. No background processes, no network calls — just file reads and terminal commands.

Troubleshooting

"Backprop CLI is not installed" Run npm install -g @mcptoolshop/backprop and reload the window, or click "Install Now" when prompted.

No experiments in the sidebar Experiments appear after your first backprop run. The data lives in ~/.backprop/experiments.json.

GPU status shows "N/A" The GPU status bar calls nvidia-smi directly. Ensure NVIDIA drivers are installed and nvidia-smi is on your PATH.

MIT License · MCP Tool Shop

mcp-tool-shop