dasomel — OSS Workbench

Requirements

macOS 14+ (Apple Silicon)
Homebrew
colima, kubectl — brew install colima kubectl
Node 22+ / pnpm
Rust (rustup)

Apple Silicon only. On Intel Macs the MLX compute path does not exist.

Application Layout — Eight Tabs

Tab	Role
Dashboard	Live RAM/CPU monitoring, one-click start/stop of the Colima (vz) K8s cluster, MLOps stack provisioning, port-forward control
kagent Operations	Cluster AIOps — per-context kagent diagnostics, toggling AI agents (security / promql / observability), connecting to the kagent UI (8090)
Pipeline	Card-based visualization of each stage: cluster start → provisioning → model download → fine-tuning → MLflow registration → serving
Model Hub	One-click flow from Hugging Face model search → host download → SeaweedFS S3 upload → MLflow Model Registry registration, plus a registered-model listing
MLX Studio	Installs the host MLX venv, runs LoRA fine-tuning on local models (live progress/loss), starts and stops `mlx_lm.server` model serving
Data	Data ingestion DAG pipeline (web/file/HF → chunking → LanceDB RAG → SeaweedFS S3 backup), DVC dataset version management
Access Console	One-click credential-free access to provisioned services (MLflow, SeaweedFS Filer, …) with health status
Air-Gap Management	Downloads offline bundles (images, charts, binaries) for closed networks, performs offline installation, verifies asset versions

Running It

1. Install dependencies

pnpm install

2. Development mode

beforeDevCommand starts the vite dev server automatically.

pnpm tauri dev

3. Start the cluster

Pressing Start Cluster on the Dashboard tab runs the following internally, with CPU/memory values computed automatically from the detected host RAM.

colima start --cpu <N> --memory <M> --vm-type=vz --mount-type=virtiofs --kubernetes

colima is not reentrant — run one lifecycle operation at a time, and never above the computed profile.

4. Provision the MLOps stack

Pressing Provision MLOps Stack applies the MLflow / SeaweedFS (+ credential Secret) / mac-gpu-bridge manifests to the cluster.

5. Port forwarding

Pressing Start Port Forwarding exposes the following addresses.

MLflow: http://127.0.0.1:5001
SeaweedFS S3 API: http://127.0.0.1:8333
SeaweedFS Filer UI: http://127.0.0.1:8888

Port forwards die with their parent process. If access drops suddenly, first check whether the process owning the forward is still alive.

6. Download → fine-tune → serve

Download a model on the Model Hub tab, then run fine-tuning and serving on the MLX Studio tab. The Pipeline tab shows the end-to-end flow; the Access Console tab covers service access.

Measured Performance (reference)

Measured on Apple M4 Pro / 64GB through the packaged app, 2026-07-27 to 07-28. Numbers vary with model, prompt, and hardware.

Item	Measured	Conditions
VLM serving throughput	196–198 tok/s (server-reported)	Qwen2-VL-2B-Instruct-4bit, mlx-vlm 0.6.7, OCR request with image
VLM serving TTFT	442–767 ms	same as above
LoRA fine-tuning (with vision stack)	674.5M trainable params (30.5%), 8.7GB peak memory	Qwen2-VL-2B bf16, `--train-vision`
K8s VM overhead	auto-computed from host RAM (64GB host → 12GB/6CPU VM)	compute runs on the host, outside the VM

Command-line Gates

make help is the entrypoint and its recipes are canonical. Two gates are worth knowing by name.

make verify          # tests + clippy + tsc + design lint + web build
make verify-airgap   # offline-startup probe

Build / Packaging

pnpm tauri build   # produces .app / .dmg bundles (unsigned local build)

Outputs:

src-tauri/target/release/bundle/macos/KubeMetal.app
src-tauri/target/release/bundle/dmg/KubeMetal_0.1.0_aarch64.dmg

If a valid codesigning identity exists in the keychain, make app signs with it automatically. A packaged app that reaches LAN clusters needs stable code signing — see External Cluster Integration for why.

Troubleshooting

When you doubt what the app reports, check directly from the CLI.

colima status --json
kubectl --context colima get pods -n default
 
# external cluster
kubectl --context <context> get pods -n kubemetal