Run Claude Code, OpenCode & Frontier Coding Models on Your Own AI Infrastructure with DEH

A practical guide to running Anthropic's Claude Code and the open-source OpenCode terminal agents with frontier open-weight models served on-premises from Dell Enterprise Hub — with steps presented side by side so you can pick one or run both.

---

TL;DR

• Build AI on premises with Dell Enterprise Hub (DEH) from Dell Technologies + Hugging Face.
• Claude Code (Anthropic) and OpenCode (open source) both point straight at that DEH endpoint — just set a base URL and a model name. No translation layer or proxy is needed.
• The payoff: a fully air-gapped, data-sovereign coding agent running a frontier open source model on your own Dell PowerEdge platforms — no tokens leave your data center.

---

How to use this guide: Every command lives in its own code block so you can copy it straight to your terminal. Replace INFERENCE_HOST with your host where your model is deployed.

Why

Pairing a best-in-class agent CLI with open-weight frontier models on your own Dell hardware gives you:

• Data sovereignty — inference happens on local GPUs, behind your firewall.
• Model control — pin a version, fine-tune it, audit it.
• Capacity-based cost — pay for GPUs, not per token.
• Predictable latency — no third-party rate limits.

DEH serves models with vLLM / SGLang, exposing the standard OpenAI REST API (/v1/chat/completions, /v1/models). It also publishes "Goodput" scenarios with per-GPU SLOs so you can size a deployment:

Agentic coding pushes large contexts, so Long context or Balanced on H200/B300 is a strong default for both agents.

---

Frontier open-source coding models on Dell Enterprise Hub

Here are the frontier open models best suited to coding / agentic software engineering.

Purpose-built / agentic coding models

Strong general models with excellent coding ability

Quantization: Models ship in BF16 / FP8 / NVFP4. NVFP4 (4-bit, for Blackwell B300/GB10) shrinks memory and boosts throughput at minimal quality loss — ideal for fitting big coding MoEs on fewer GPUs.

How to pick

• Single workstation (Dell Pro Max GB10): GPT-OSS-20B, Qwen3.5-27B, or quantized Qwen3-Coder-Next.
• Team server (XE9680 / H100 / H200): Qwen3-Coder-Next (80B) — the coding-tuned, Apache-2.0 sweet spot.
• Max capability (XE9780 / B300): GLM 5.1, Kimi K2.6, or DeepSeek V4 Pro for long-horizon, whole-repo work.

---

Step-by-step

Both agents follow the same five steps: deploy a model → install the CLI → point it at the endpoint → run. The only differences are the config file each agent uses.

Step 1 — Deploy a model on Dell Enterprise Hub (shared)

1. Log in to https://dell.hf.co.
2. In the Model Catalog, filter by your Dell Platform and pick a coding model — e.g. Qwen3-Coder-Next.
3. Click Deploy, choose the platform and a Goodput scenario (Balanced or Long context).
4. Run the generated container command on your Dell PowerEdge server. It serves an OpenAI-compatible API on port 8000.

Check that the endpoint is live:

curl http://INFERENCE_HOST:8000/v1/models

Both Claude Code and OpenCode will talk to this same endpoint. No proxy, no gateway, no translation layer.

Step 2 — Install the agent

Claude Code (native installer, no Node.js needed):

curl -fsSL https://claude.ai/install.sh | bash
claude --version

OpenCode (install script):

curl -fsSL https://opencode.ai/install | bash
opencode --version

Step 3 — Point the agent at your DEH model

Pick the tab for your agent and paste the block. Both write a single config file. Replace INFERENCE_HOST and the model name.

Claude Code — ~/.claude/settings.json:

mkdir -p ~/.claude && cat > ~/.claude/settings.json <<'EOF'
{
  "env": {
    "ANTHROPIC_BASE_URL": "http://INFERENCE_HOST:8000",
    "ANTHROPIC_API_KEY": "dummy",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3-coder-next"
  },
  "theme": "dark"
}
EOF

• ANTHROPIC_BASE_URL → your DEH endpoint instead of api.anthropic.com.
• ANTHROPIC_API_KEY → dummy (a local endpoint ignores it; use a real key for a secured endpoint).
• ANTHROPIC_DEFAULT_*_MODEL → maps Claude's three tiers onto your model. Point them at different DEH models if you want a big model for Opus and a fast one for Haiku.

OpenCode — ~/.config/opencode/opencode.json:

mkdir -p ~/.config/opencode && cat > ~/.config/opencode/opencode.json <<'EOF'
{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "vllm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Dell Enterprise Hub (local)",
      "options": { "baseURL": "http://INFERENCE_HOST:8000/v1" },
      "models": { "qwen3-coder-next": { "name": "Qwen3 Coder Next" } }
    }
  },
  "model": "vllm/qwen3-coder-next"
}
EOF

• options.baseURL → the DEH endpoint, including /v1.
• model → the active model as provider/model.

Step 4 — Add credentials (OpenCode only)

Claude Code already has its key in settings.json. OpenCode stores credentials separately:

mkdir -p ~/.local/share/opencode && cat > ~/.local/share/opencode/auth.json <<'EOF'
{ "vllm": { "type": "api", "key": "dummy" } }
EOF

Use a real key for a secured endpoint, and never commit auth.json.

Step 5 — (Optional) Add project conventions

Each agent reads a guidance file from your project root each session — CLAUDE.md for Claude Code, AGENTS.md for OpenCode. Create both at once:

cat > CLAUDE.md <<'EOF'
# Project conventions

## Python environment
This project uses **uv**. Do NOT use `pip` directly.
- Run a script: `uv run python <script>`
- Run tests:    `uv run pytest`
- Add a dep:    `uv add <package>`
- Sync env:     `uv sync`
EOF
cp CLAUDE.md AGENTS.md

Tip: in OpenCode you can run /init to auto-generate AGENTS.md from your repo.

Step 6 — Run it

cd ~/my-project
claude        # Claude Code
# or
opencode      # OpenCode

Every request now flows straight to your DEH model on your own hardware.

---

Config & file-location cheat sheet

---

Validation & troubleshooting

Quick endpoint sanity check (bypasses both agents):

curl http://INFERENCE_HOST:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-coder-next","messages":[{"role":"user","content":"Reverse a linked list in Rust."}]}'

---

Production hardening checklist (both)

• Pin the model version in DEH for reproducibility.
• Right-size with Goodput SLOs to match context + concurrency.
• Tier models: large for primary (GLM 5.1 / Kimi K2.6), fast for auxiliary (GPT-OSS-20B).
• Secure shared endpoints with TLS + real keys; never commit secrets.
• Prefer NVFP4/FP8 on Blackwell (B300/GB10) to fit bigger coding MoEs per GPU.
• Keep CLAUDE.md / AGENTS.md rich — explicit build/test/lint commands improve open-model reliability.

---

7. Which should you choose?

• Want the most polished, batteries-included agent UX? Claude Code.
• Want a 100% open stack? OpenCode.
• They use entirely separate config directories, so you can run both side by side in the same repo and cross check each other's work.

---

Conclusion

Whether you choose Anthropic's Claude Code or the open-source OpenCode or both, the destination is the same: a frontier open-weight coding model running on your own Dell PowerEdge Platforms, with your source code never leaving the data center.

You can start with Qwen3-Coder-Next or gemma-4-31B-it, then scale up to GLM 5.1 / Kimi K2.6 / MiniMax2.7 / DeepSeek V4 / Nemotron 3 Ultra on H200/B300 for the most demanding work. Either harness, any recommended frontier coding model and all on-prem.

---

References

• Dell Enterprise Hub - https://dell.hf.co (Model Catalog, App Catalog, Docs, Optimized Deployments, Security, Goodput Scenarios)
• Anthropic Claude Code - Build, debug, and ship with natural language - https://claude.com/product/claude-code
• OpenCode - The open source AI coding agent - https://opencode.ai/