ollama-local_skill

Overview

This skill manages and uses local Ollama models for inference, embeddings, and integrated tool use. It provides CLI and script helpers for listing, pulling, removing, showing models, chatting, generating completions, and producing embeddings. It also supports OpenClaw sub-agent spawning and multi-agent patterns for collaborative tasks. Use it to run local LLM workflows without cloud dependencies.

How this skill works

The skill talks to an Ollama daemon (default http://localhost:11434) via environment-configured host or direct API calls. It wraps common operations—list, pull, remove, show models—and exposes chat, generate, and embed commands plus a tool loop for function-calling models. For multi-agent or persistent sessions, it spawns OpenClaw sub-agents using model paths like ollama/<model-name> and coordinates tool-enabled interactions.

When to use it

• You need local inference or embeddings without relying on cloud APIs.
• Managing model lifecycle (list, pull, remove, inspect) on a local Ollama host.
• Running tool-enabled conversations or function-calling workflows locally.
• Spawning sub-agents for parallel or collaborative tasks (think tank pattern).
• Integrating Ollama models into custom scripts or CI pipelines via direct API calls.

Best practices

• Set OLLAMA_HOST environment variable to point to your Ollama server before running commands.
• Start with smaller models (7b, 4b) for development; move to larger models only after verifying VRAM and performance.
• Use the tool loop for workflows that require model-driven tool invocation and iterative calls.
• Verify firewall and port 11434 access for remote Ollama hosts and ensure ollama serve is running.
• Use explicit model tags (e.g., qwen3:4b) and the ollama/<model-name> path when spawning OpenClaw sub-agents.

Example use cases

• Pull and run a local 8b model for offline chat and code review tasks.
• Generate embeddings from documents for a private semantic search index.
• Spawn three specialized sub-agents (architect, coder, reviewer) to collaboratively design and implement a feature.
• Use the tool-enabled loop to let a model search, fetch results, and synthesize a summary.
• Integrate Ollama chat or generate endpoints into a local web app via simple curl or script calls.