AiAuN
# AiAuN
Local-light **MCP** orchestrator: GitHub for code, Kaggle for data + GPU, W&B for runs, Google Drive for small artifacts. The laptop only wraps APIs; training happens on Kaggle.
## Docs
| Doc | What |
|-----|------|
| [docs/VISION.md](docs/VISION.md) | Product intent |
| [docs/SETUP.md](docs/SETUP.md) | Keys, WSL install (`mcp` package), Cursor MCP, add to agents |
| [docs/USAGE.md](docs/USAGE.md) | Prompts, tools, features |
| [docs/design.md](docs/design.md) | Architecture and API flow |
| [docs/implementation.md](docs/implementation.md) | Package layout and how to extend |
## Quick start
1. Fill `.env` from `.env.example` (see Setup).
2. WSL: `bash scripts/install_wsl.sh` (needs Python ≥3.10 + official `mcp`; tries `ssemi` then `base` then `uv` / conda `aiaun`).
3. Copy [`.cursor/mcp.json.example`](.cursor/mcp.json.example) to `.cursor/mcp.json` and enable the `aiaun` server in Cursor.
4. Ask the agent to run an experiment; it should **ask** for data dir, code version, and config if those are missing.
Phase-1 smoke uses `fixtures/synthetic_color_cls` (tiny PPMs), not COCO re-upload.
TDQS
Scored across 14 tools
Each tool targets a distinct step in the Kaggle/Drive experiment workflow: validation, local inspection, dataset/kernel operations, script source, tracking, and Drive transfer. Only the two smoke-script tools are close variants, but their names and descriptions clearly separate color-class vs. ResNet50 GPU usage.
Names are readable but grammatically inconsistent: some are verb-first like inspect_local_dir, some are object-verb like kaggle_kernel_push, and others are noun phrases like experiment_tracking_links or drive_folder_info. The kaggle_* and drive_* prefixes provide some grouping, but there is no single predictable verb_noun convention.
14 tools is a reasonable size for an experiment pipeline spanning local inspection, Kaggle dataset/kernel lifecycle, and Drive output transfer. Each tool maps to a needed step, so the count feels well-scoped.
The set covers the main loop: validate request, inspect data, push dataset, push kernel, poll logs, move outputs to Drive, and provide tracking links. Minor gaps exist—configs can be listed but not read, and there is no update/cancel for kernels—but agents can generally work around them.