Coder Cloud Development Environments

Interactive Cloud Development for Berkelium

Interactive IDEs & AI Workspaces, Powered by Kubernetes

Coder provides a seamless, web-based and desktop IDE experience on top of the Berkelium cluster. It complements Berkelium's native Kubernetes API by giving researchers instant access to VS Code, Zed, JupyterLab, autonomous AI agents, and our fleet of 32 GPUs (NVIDIA H100, A100, A40, AMD MI100) without requiring any Kubernetes configuration or YAML files.

Launch Coder PortalRead Coder Documentation
VS Code & JetBrains
Browser or Local Desktop
NVIDIA H100 / A100 / A40
Instant GPU Allocation
AMD Instinct MI100
ROCm Accelerated Compute
AI Coding Agents
Claude Code, OpenClaw, Cursor

Two Complementary Ways to Compute on Berkelium

Science IT provides both Direct Kubernetes Access (via kubectl, Helm, and KubeRay) for scalable production pipelines, and Coder Workspaces for interactive development and rapid prototyping. Both share the same underlying 8 compute nodes, CephFS data fabric, and security controls.

Learn how they integrate
1-Click
Instant Provisioning
32 GPUs
H100, A100, A40 & MI100
LBNL OneID
Institutional Single Sign-On
100% On-Cluster
Zero Data Egress Risk

Capabilities Designed for Research Velocity

Engineered to streamline daily interactive work so computational researchers, postdocs, and PIs can prototype and build without friction.

Any IDE, Anywhere

Access full-featured VS Code in your web browser, or connect your local desktop VS Code or Zed IDEs via Coder SSH Gateway.

  • Web-based Code Server with terminal
  • Local desktop IDE SSH bridge
  • Integrated JupyterLab & RStudio

AI Coding Agents & Governance

Run modern AI pair programmers and autonomous agents—including Claude Code, OpenCode, Aider, OpenClaw, Cursor, and local vLLM / Ollama models—pre-wired to LBNL CBorg.

  • Audit logging for AI tool calls
  • Secure API token injection via Vault
  • Support for in-cluster local LLM endpoints

NVIDIA & AMD Acceleration

Attach dedicated NVIDIA H100 (80GB HBM3), A100 (80GB), A40 (48GB), or AMD Instinct MI100 (32GB) accelerators directly to your workspace with a single dropdown.

  • Pre-installed CUDA, cuDNN, ROCm & PyTorch
  • Interactive model debugging & training
  • Zero Slurm batch queue wait times

16 Production Lab Templates

Ensure scientific reproducibility across your research group. Spin up pre-configured templates for PyTorch/CUDA, AI agents, vLLM, ComfyUI, Marimo, bioinformatics, RStudio, Selkies KDE & XFCE Linux GUI desktops, and KubeVirt VMs.

  • 16 curated AI, GPU, scientific, desktop & VM templates
  • Auto-clone Git repos & personal dotfiles
  • Pre-installed uv, ripgrep, gh & nvtop

Persistent Scientific Storage

Your home directories, datasets, and virtual environments persist seamlessly across workspace restarts via high-speed Ceph RBD / CephFS and local NVMe scratch volumes.

  • Persistent Ceph RBD /home/coder volumes
  • In-memory /dev/shm for multi-GPU PyTorch
  • Dedicated /scratch block disks for KubeVirt VMs

DOE & Institutional Security

Fully self-hosted in the LBNL data center with LBNL OneID SSO authentication. Code and sensitive research datasets never leave the lab boundary or go to public clouds.

  • Zero-trust network policy isolation
  • Automatic session auto-stop & quota caps
  • LBNL division & group-based RBAC
16 Pre-Configured Environments

Available Workspace Templates

Every template is pinned to an immutable release image and includes persistent Ceph RBD storage, modern CLI tooling (uv,ripgrep, fd, fzf,gh, tmux, btop), and 1-click browser or desktop IDE access.

AI Agents, LLM Inference & Generative Media

AI Coding Agents Studioai-agents

Claude Code, OpenCode, Aider, Gemini CLI, Codex CLI & Continue pre-wired to LBNL CBorg or NRP LLMs.

Apps: VS Code Web (:13337) · Live Preview (:3000) · Cursor
OpenClaw Agent Gatewayopenclaw

Autonomous OpenClaw AI assistant with Web Control UI, live LLM provider sync (CBorg / NRP), and Node 24 LTS.

Apps: OpenClaw UI (:18789) · VS Code Web (:13337) · Cursor
vLLM Inference & Playgroundvllm · 1–8 GPUs

High-throughput vLLM OpenAI server with multi-GPU tensor parallelism, hf_transfer, and built-in Gradio benchmark UI.

Apps: Chat Playground (:7860) · OpenAI API (:8000) · VS Code
Ollama + Open WebUIollama-webui · 0–4 GPUs

1-click private local LLM & RAG workspace with automatic starter model pre-pull and persistent Open WebUI vector store.

Apps: Open WebUI (:8080) · Ollama API (:11434) · VS Code
ComfyUI Diffusion Studiocomfyui · 1–4 GPUs

GPU-accelerated generative image, video & audio workflow studio with ComfyUI, ComfyUI-Manager, and JupyterLab.

Apps: ComfyUI (:8188) · JupyterLab (:8888) · VS Code (:13337)

GPU Compute, Data Science & Scientific Research

PyTorch / CUDAcuda · 0–8 GPUs

Deep learning suite with PyTorch (2.14.1+cu130 for B300 Blackwell), JupyterLab, TensorBoard, uv, andnvtop.

Apps: JupyterLab (:8888) · TensorBoard (:6006) · VS Code
Reactive Data Labdata-lab · 0–4 GPUs

Marimo reactive notebooks, DuckDB OLAP SQL, Polars, PyArrow, JupyterLab, and live Streamlit / Gradio app previews.

Apps: Marimo (:2718) · Jupyter (:8888) · Streamlit (:8501)
Bio & Chem Labbio-chem-lab · 0–4 GPUs

Bioinformatics & molecular modeling with Biopython, RDKit, MDAnalysis, py3Dmol, Scanpy, AnnData, and PyTorch.

Apps: JupyterLab (:8888) · VS Code Web (:13337) · Cursor
RStudio Serverrstudio

Statistical computing and biostatistics environment with RStudio Server, Tidyverse, and VS Code R extensions.

Apps: RStudio (:8787) · VS Code Web (:13337) · Cursor

Virtual Machines, GUI Desktop & Systems Engineering

KubeVirt Linux VMkubevirt · Full KVM VM

Hardware-virtualized Ubuntu 24.04/22.04 or Debian 12/13 VM withsystemd, Docker Engine + Compose, and optional/scratch block disk.

Access: Web Terminal · coder ssh · Optional VS Code
Selkies GPU Desktop (KDE)selkies · 0–4 GPUs

Hardware-accelerated KDE Plasma 6 desktop via Selkies (EGL/NVENC, PipeWire audio, X11/Wayland) with PyCharm, Blender, and Wine.

Apps: KDE Desktop (:8080) · VS Code Web (:13337) · Cursor
Linux GUI Desktop (XFCE)desktop · 0–4 GPUs

Browser HTML5 noVNC XFCE4 desktop with Minimal, Scientific (Blender, ParaView, GIMP), or Engineering (KiCad, QGIS, Wireshark) bundles.

Apps: XFCE Desktop (:6080) · VS Code Web (:13337) · Cursor
Cloud-Native (Go + Rust)cloud-native

Systems & Kubernetes engineering with Go, Rust (rustup, cargo,rust-analyzer), kubectl,helm, k9s, and talosctl.

Apps: VS Code Web (:13337) · Cursor · Web Terminal
Go (Golang)golang

Dedicated Go workspace with persistent GOPATH,gopls, Delve debugger (dlv),kubectl, and helm.

Apps: VS Code Web (:13337) · Cursor · Web Terminal
Node.js 24 LTS & Webnodejs

Modern TypeScript & full-stack web studio with Node 24 LTS,pnpm, yarn, bun,tsx, and NAT64 IPv6 preload.

Apps: VS Code Web (:13337) · Web Preview (:3000) · Cursor
Universal Polyglotuniversal

Multi-language Ubuntu environment with Python, Node.js, Go, Java, Ruby, C/C++, uv, and modern CLI utilities.

Apps: VS Code Web (:13337) · Cursor · Web Terminal

Choosing the Right Workflow for Your Science

Both workflows operate on the same Berkelium cluster. Choose the interface that best fits your immediate computational task:

Use Coder Workspaces For

Interactive Science, AI Agents, GUI Desktops & VMs
  • Interactive Notebooks & Analytics: Exploratory Python, R, and SQL analysis in JupyterLab, Marimo reactive notebooks, RStudio, and DuckDB
  • AI-Assisted Coding & Autonomous Agents: Pair programming with Claude Code, OpenCode, Aider, OpenClaw, and Cursor pre-wired to LBNL CBorg
  • Turnkey LLM Serving & Generative Studios: 1-click vLLM multi-GPU playgrounds, private Ollama + Open WebUI RAG stacks, and ComfyUI diffusion workflows
  • Iterative GPU & Scientific Computing: Interactive PyTorch debugging, TensorBoard profiling, and molecular modeling (RDKit, MDAnalysis, Biopython) on NVIDIA B300, H200, H100, A100, and A40 GPUs
  • Linux GUI Desktops & Full KVM VMs: Browser-streamed Selkies KDE Plasma & XFCE4 desktops (ParaView, Blender, QGIS, KiCad) and isolated KubeVirt Linux VMs with Docker Engine
  • Zero-Setup Onboarding: Getting students, postdocs, and collaborators coding in VS Code Web, Go/Rust toolchains, or scientific stacks in under 60 seconds

Use Native Kubernetes For

Distributed Compute, Batch Pipelines & Scaled Services
  • Multi-Node Distributed Training: Large-scale PyTorch (DDP / FSDP / DeepSpeed) and MPI jobs spanning multiple GPU nodes
  • Elastic Ray Clusters: Distributed hyperparameter tuning, reinforcement learning, and data processing via the KubeRay operator
  • Production Web Portals & APIs: High-availability web services, databases, and microservices exposed via cluster Ingress
  • Unattended Batch Pipelines: Large-scale automated Kubernetes Jobs, CronJobs, and high-throughput parameter sweeps
  • Custom Cloud-Native Infrastructure: Helm deployments, StatefulSets, custom operators, and GitOps / CI/CD automation
  • Multi-Pod Service Topologies: Coupled simulation, message-queue, and telemetry stacks orchestrated via Kubernetes manifests

Get Started in 3 Minutes

1

Sign In via LBNL SSO

Navigate to coder.bk.lbl.gov and log in using your institutional Berkeley Lab OneID credentials.

2

Choose Your Environment

Select a workspace template suited for your workflow and configure your CPU, memory, and GPU resources.

3

Launch & Start Coding

Open VS Code directly in your browser or connect through your local desktop IDE using the Coder CLI and JetBrains Gateway.

Ready to elevate your scientific workflow?

Launch a Coder workspace today on Berkelium, or explore how to scale into native Kubernetes batch jobs and Ray clusters in our documentation suite.

Launch Coder Portal 🚀Explore Coder Documentation →