A desktop research IDE for running, tracking, and visualizing ML experiments. Built with Electrobun + React + Tailwind.
Llama Lab gives you a single place to manage research programs — from experiment code to results to dependency graphs. It's built for the workflow of iterating on ML training scripts: edit code, run experiments, compare results, spawn AI agents to help, repeat.
Programs & Experiments — Define research programs with tracks, metrics, and dependencies. View experiment results with sortable tables and metadata.
Source Tree & Editor — Browse and edit your codebase directly. Syntax-aware monospace editor with Cmd+S save.
Git Integration — Full source control panel: status, staging, diff viewer, commit, push, branch switching. All without leaving the app.
Dependency Graph — Interactive node graph showing how research programs relate to each other. Auto-layout with drag-to-rearrange.
Experiment Charts — SVG charting for CSV step logs (loss curves, training metrics) and TSV/JSON summary data (bar charts for comparing experiments). Auto-discovers log files and detects schemas from headers.
Multi-Session Terminal — Multiple named terminal tabs with full PTY support. Open terminals scoped to specific program directories.
AI Agent Spawning — Launch autonomous agents (Claude, Codex, OpenClaw, Hermes) scoped to a research program with contextualized prompts. Agents run in tmux sessions and survive app restarts. Live log tailing.
Command Palette — Cmd+P to search programs, files, and commands.
Workspace Management — First-run setup wizard, multiple workspaces, recent workspace switching.
# Clone
git clone https://github.com/kortexa-ai/llamalab.desktop.git
cd llamalab.desktop
# Install and run (picks the right mode automatically)
./run.sh
# Or explicitly:
./run.sh hmr # Vite HMR + Electrobun (recommended for dev)
./run.sh dev # Vite build + Electrobun watch (no HMR)
./run.sh build # Production build
./run.sh clean # Nuke cachesRequires Bun and Electrobun.
On first launch, a setup wizard will guide you through configuring your workspace and connecting a source code repository.
src/
bun/ # Backend (runs in Bun/Electrobun main process)
index.ts # Entry point, RPC wiring, window/menu setup
config.ts # Centralized workspace config (single source of truth for paths)
research.ts # Program/track/results data layer (filesystem reads)
git.ts # Git CLI wrapper (status, diff, commit, push, branches)
charts.ts # Experiment log discovery, CSV/TSV/JSON parsing
agents.ts # AI agent spawn/list/kill via tmux
prompt-builder.ts # Contextualized prompt generation for agents
terminal.ts # PTY session manager (Python wrapper)
workspace.ts # Workspace CRUD, setup wizard backend
preferences.ts # App preferences (~/.config/llamalab/)
mainview/ # Frontend (React, renders in Electrobun webview)
App.tsx # Root layout, setup wizard gate
hooks/
useWorkspace.tsx # Global state (React context + reducer)
components/
Sidebar.tsx # Programs, source tree, source control, agents
ContentPane.tsx # Tab content router
ProgramOverview.tsx
GraphView.tsx # Interactive dependency graph (@xyflow/react)
ChartView.tsx # SVG line/bar charts + charts browser
FileEditor.tsx # Code editor with save
DiffViewer.tsx # Unified diff with color
TerminalPanel.tsx # Multi-session terminal (@xterm/xterm)
CommandPalette.tsx
SetupWizard.tsx # First-run / new workspace wizard
...
rpc.ts # Electroview RPC setup, event bridging
shared/
types.ts # All RPC types, data interfaces
Communication between backend and frontend uses Electrobun's typed RPC system. The full schema lives in src/shared/types.ts.
- Runtime: Bun + Electrobun (macOS native, no Electron)
- Frontend: React 18, Tailwind CSS, TypeScript
- Graph: @xyflow/react
- Terminal: @xterm/xterm + Python PTY wrapper
- Icons: @phosphor-icons/react
- Charts: Pure SVG (no charting library)
- Build: Vite (HMR in dev)
See CONTRIBUTING.md. Issues and PRs welcome.
MIT