A local fleet management skill for running high-performance DeepSeek models across multiple devices without cloud dependencies.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install deepseek-deepseek-v3
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install deepseek-deepseek-v3 using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is DeepSeek Local Fleet?
The DeepSeek skill enables developers to run a distributed fleet of DeepSeek models, including V3, R1, and Coder variants, on their own hardware using Ollama Herd. By utilizing Openclaw Skills to manage these local deployments, users can eliminate per-token cloud costs while ensuring that sensitive code and business data never leave the local network. This setup transforms a collection of machines, such as Apple Silicon Macs or Linux servers, into a unified inference engine.
This skill provides an intelligent routing layer that automatically picks the best machine for every request based on 7-signal scoring. It handles the complexities of model distribution, allowing you to access powerful reasoning and coding capabilities through a single, OpenAI-compatible API endpoint.
DeepSeek Local Fleet Use Cases
- Running private, large-scale reasoning tasks using DeepSeek-R1 without sharing data with cloud providers.
- Optimizing local code generation workflows with DeepSeek-Coder-V2 across a network of Mac devices.
- Building cost-effective AI agents that leverage high-parameter models like DeepSeek-V3 on a local homelab fleet.
- Scaling inference capacity by pooling VRAM from multiple local machines into a single unified endpoint.
- Bypassing cloud API rate limits and outages by maintaining local model availability.
How DeepSeek Local Fleet Works
- Install and initialize the herd router on a primary controller to manage incoming traffic and model distribution.
- Connect individual machines as nodes to the fleet using the node command to contribute their local CPU and GPU resources.
- Pull the desired DeepSeek models via Ollama on the respective nodes according to their hardware capacity.
- The 7-signal scoring system analyzes connected nodes to identify the most efficient machine for a specific model request.
- Requests are automatically routed to the optimal node, with transparent failover and VRAM-aware fallback if a machine is busy.
- Performance and model health are monitored through an integrated dashboard, providing real-time visibility into the fleet.
DeepSeek Local Fleet Setup
To begin using this capability within your Openclaw Skills collection, install the fleet manager and start the router across your network:
pip install ollama-herd
herd # Start the router on port 11435
herd-node # Run on each contributing machine
# Pull your preferred DeepSeek model
ollama pull deepseek-r1:70b
DeepSeek Local Fleet Data Schema & Taxonomy
The skill maintains its state and performance history in a structured format within the local user directory to optimize routing decisions within the Openclaw Skills ecosystem:
| Component | Path | Description |
|---|---|---|
| Latency DB | ~/.fleet-manager/latency.db |
SQLite database tracking model performance and node response times |
| Request Logs | ~/.fleet-manager/logs/herd.jsonl |
Structured JSONL logs of all requests, routing decisions, and token throughput |
| Configuration | Local Environment | Environment variables for port mapping and node discovery |
DeepSeek Local Fleet Advanced Features
- 7-signal scoring for real-time optimal node selection based on latency and load.
- VRAM-aware fallback to prevent cold-loading models when a suitable alternative is already active in memory.
- Context protection mechanisms to avoid expensive model reloads caused by changes in the context window.
- Multi-modal support including integrated image generation and speech-to-text endpoints on the same fleet.
- Automatic retry logic that fails over to the next best node transparently if a request fails.
SKILL.md
Loading
Related Openclaw Skills
Ollama Herd
A high-performance multimodal router that unifies multiple local Ollama instances into a single, smart API endpoint.
Ville de Québec Open Data
A specialized interface to access, filter, and export municipal datasets from the City of Quebec directly through your AI agent.
Automated TikTok & Instagram Trend Radar
A fully automated trend monitoring and video production pipeline that detects viral social media spikes and generates ready-to-publish content.
Social Media Data Extractor
A multi-platform social media intelligence engine that scrapes, normalizes, and analyzes engagement across five major platforms using Apify and Claude AI.
Spec-Kit
Spec-Kit is a framework for Spec-Driven Development that turns specifications into first-class citizens for AI agent workflows.
Twitter Command Center (Search + Post + Interact)
An all-in-one Twitter intelligence and automation tool for autonomous agents to search, monitor, and interact with X data.
Featured*
