A personal education project born from wanting to understand software and AI models at a deeper level. What started as a learning exercise turned into something I genuinely didn't expect.
BigEd CC is a centralized model manager toolkit that lets you control local LLMs (via Ollama), API-based LLMs (Claude, Gemini), and OAuth models (via VS Code with pre-configured files and a first-time walkthrough). It grew out of curiosity and vibe-coding sessions into a 130+ skill autonomous agent fleet — and I'm still figuring out what it all means.
This is not a product. It's a reference project for anyone curious about multi-model orchestration, agent lifecycle patterns, or what happens when you let AI help build the thing that manages AI.
- Model Management — Switch between local (Ollama), Claude, and Gemini from one interface. See which model is active, what it costs, and how fast it responds.
- 130+ AI Skills — Code review, security audit, web research, ML training, knowledge indexing, and more — all dispatched to agents automatically.
- Dynamic Agent Scaling — 4 core agents + demand-based scaling up to 16, based on your hardware.
- Dr. Ders — A hardware supervisor that monitors thermals, VRAM, and model health so things don't melt.
- Fleet Dashboard — Real-time web UI at localhost:5555 showing agent status, task queues, and performance metrics.
- One-Click Setup — Installer handles Python, Ollama, models, and dependencies. There's also a first-run walkthrough in the GUI.
- Auto-Save Backup — Periodic snapshots so you don't lose fleet state.
- Cost Tracking — Per-call token usage and estimated costs for API providers.
Nearly the entire codebase was generated through AI-assisted development. That's part of the point — this project is itself a case study in what vibe-coding produces at scale.
| Model | Role | Estimated Contribution |
|---|---|---|
| Claude Code (Opus 4.6) | Primary architect — wrote most of the code, designed fleet architecture, built the skill system | ~70% |
| Claude (Sonnet 4.6) | Code review, audits, skill generation, iterative improvements | ~15% |
| Gemini Pro (2.5/3.1) | Independent reviews, architecture audits, second opinions | ~10% |
| Human (Max) | Direction, judgment calls, testing, and the occasional manual fix when models got stuck | ~5% |
Download Setup.exe from Releases → Run → Follow wizard
git clone https://github.com/mbachaud/BigEd.git cd BigEd python fleet/dependency_check.py # pre-flight check python fleet/smoke_test.py --fast # verify 51/52 smoke tests python BigEd/launcher/launcher.py # launch GUI
BigEd CC
├── BigEd/launcher/ — GUI launcher (PyWebView + Flask dashboard)
│ ├── modules/ — Pluggable modules (Intelligence, Ingestion, Outputs)
│ └── fonts/ — Custom pixel fonts
├── fleet/ — 130+ skill AI worker fleet
│ ├── supervisor.py — Process lifecycle + dynamic scaling
│ ├── hw_supervisor.py — Dr. Ders (thermal + model management)
│ ├── dashboard.py — Web dashboard (localhost:5555)
│ ├── worker.py — Generic task executor
│ ├── skills/ — 130+ registered skills
│ └── knowledge/ — Agent-generated artifacts
├── autoresearch/ — ML training pipeline (inspired by Karpathy)
├── scripts/ — Setup scripts (Windows/Linux/macOS)
└── docs/ — Specs, flowcharts, design docs
| Provider | Models | Auth | Cost |
|---|---|---|---|
| Ollama (Local) | qwen3:8b, 4b, 1.7b, 0.6b | None | Free |
| Claude | Haiku, Sonnet, Opus | API key or OAuth | Per-token |
| Gemini | Flash, Pro | API key or OAuth | Per-token |
OAuth models (Claude Code, Gemini CLI) work through VS Code with pre-configured project files — BigEd writes a task briefing, opens VS Code, and you're ready to go.
Because the models kept building, BigEd ended up with features I didn't originally plan for:
- File access control — Per-zone permissions (read / read-write / full)
- DLP — Secret detection and output scrubbing before API calls
- Audit logging — API calls, file access, and config changes are logged
- Air-gap mode — Full offline operation with local models only
- Fleet federation — Multi-device task routing (experimental, in
fleet.toml) - Training pipeline — Autonomous ML training loop inspired by Karpathy's build-nanogpt
These work, but they haven't been battle-tested beyond my own machine. Take them for what they are.
| Repo | Purpose |
|---|---|
| BigEd | Core platform — launcher, fleet, dashboard, skills, ML pipeline |
| BigEd-ModuleHub | Optional modules — UI extensions loaded at runtime |
BigEd ships with .vscode/launch.json and .vscode/tasks.json for shared dev
configs (debug launchers, smoke test tasks). These are tracked in git.
MCP server configuration is not tracked — it lives in .vscode/mcp.json
and is gitignored because it may contain credentials or machine-specific
server URLs. On first clone you won't have one, so either:
Option A — Generate a starter config:
python fleet/mcp_manager.py --init-vscode
Option B — Create it manually:
{
"servers": {},
"inputs": []
}Save as .vscode/mcp.json and add your MCP servers as needed.
The root .mcp.json (used by Claude Code CLI) is also gitignored and managed
at runtime by fleet/mcp_manager.py. See OPERATIONS.md for MCP server
management commands.
See CONTRIBUTING.md for guidelines. This is a personal project, but PRs, issues, and honest feedback are welcome — especially if you spot something the models got wrong.
Apache 2.0 — see LICENSE.
Copyright 2025-2026 Michael Bachaud (mbachaud).