chestercs

Dgx Openclaw Stack — Communication skill for Claude Code

Communication community

Self-hosted OpenClaw agent stack on one NVIDIA DGX Spark / GB10, running DeepSeek-V4.1-Flash (284B MoE) on Strata: a Discord bot, a LoRa mesh bridge and offline Claude Code on one desktop box.

How to install Dgx Openclaw Stack

This entry records only its repository, not the path inside it, so there is no exact command to give. Open chestercs/dgx-openclaw-stack and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Dgx Openclaw Stack does

Self-hosted OpenClaw agent stack on one NVIDIA DGX Spark / GB10, running DeepSeek-V4.1-Flash (284B MoE) on Strata: a Discord bot, a LoRa mesh bridge and offline Claude Code on one desktop box. Nemotron on vLLM as the no-build quick start. bge-m3 memory, SearxNG, Whisper.

Alternatives in Communication

Prompts.chat — f.k.a. Awesome ChatGPT Prompts 154.3k ★
Lark Coding Agent Bridge — Bot that bridges Feishu/Lark messenger with a local Claude Code or Codex CLI 2.4k ★
X Research Skill — X/Twitter research skill for Claude Code and OpenClaw 974 ★

README

DGX OpenClaw Stack

A self-hosted [OpenClaw](https://openclaw.ai) agent on one NVIDIA GB10 box (DGX Spark / ASUS Ascent GX10), brought up with `docker compose`.

**The model: DeepSeek-V4.1-Flash, 284B parameters, on one desktop box.** It runs on [Strata](https://github.com/Niko1221/Strata)'s engine in a ~1.75-bit quantisation: the experts the model routes to most sit on the GPU, the rest is computed on the CPU straight from the files. On the reference GX10 it serves the Discord bot, the MeshCore mesh, and Claude Code on the LAN, offline. It comes back by itself after a reboot, stays below 80 °C under sustained load, and its limits allow agent runs of up to a week. It is not fast (code 20–33 tok/s, prose 15–17, one request at a time), but it is the largest model this box runs, and it runs reliably ([`openclaw-strata/README.md`](openclaw-strata/README.md)).

Around it: bge-m3 embeddings for memory, SearxNG web search, faster-whisper for voice notes, and OpenClaw's Web UI, agents and Discord bot. The quick start needs no engine build: Nemotron 3.5 Super on vLLM, one `docker compose up`; Gemma 4 and Mistral Small 4 are other vLLM options. The value is the wiring: `config-init` writes `openclaw.json` and the agent workspaces from `.env` on every `docker compose up`, so the setup stays the same across onboarding re-runs, restarts and OpenClaw upgrades.

[![License: MIT](https://img.shields.io/badge/license-MIT-yellow.svg)](LICENSE)

What runs

Service	What it does	Host port	Starts with
`strata`	Chat model: DeepSeek-V4.1-Flash (ISTA GSQ-RCO IQ1_M) on Strata's `strata-dsv41` engine, built on the host; tools, thinking, DSpark drafting, text only	`127.0.0.1:8097`	`compose/strata.yml`, started by `llm-ensure.sh` after the stack
`strata-router`	Claude Code's endpoint for it: keeps the auto-mode classifier inside its 60 s limit	`127.0.0.1:8096`	`compose/strata-router.yml`
`vllm-llm`	Quick-start chat model: Nemotron 3.5