ollama-cli

name	ollama-cli
description	Execute and manage local LLMs using the ollama CLI, including pulling models and launching agents with MCP configurations. You MUST load this skill when interacting with the `ollama` CLI.
license	MIT

ollama-cli

When to Use

• When querying, pulling, or managing local LLMs via the ollama CLI.
• When configuring local model backends for AI agents.
• When launching wrapper tools (like Claude Code) using an Ollama model with Model Context Protocol (MCP) configurations.

When Not to Use

• When making standard REST API calls to remote cloud providers (e.g., OpenAI, Anthropic) without a local proxy.
• For managing Python environments or general CLI tools unrelated to local inference.

Step-by-Step Workflows

1. Verification and Setup

1. Check Daemon Status: Run ollama list to verify the local Ollama daemon is active and responsive.
2. Pull Model: If the target model is missing, execute ollama pull <model-name> (e.g., ollama pull gemma4:e2b) to cache it locally before running.

2. Execution and Agent Launch

1. Interactive Run: Use ollama run <model-name> to test the model directly in the CLI.
2. Launch with MCP Config: When integrating Ollama models with agent workflows (like Claude Code) that require Model Context Protocol configurations, use process substitution to pass dynamic JSON configurations safely.

Command Examples

• List Available Models:

  ollama list
  

• Pull a Specific Model:

  ollama pull gemma4:e2b
  

• Launch Claude Code with Ollama and MCP Config: Demonstrates how to launch Claude via Ollama, specifying the model and injecting an inline JSON MCP configuration using bash process substitution:

  ollama launch claude --model gemma4:e2b -- --mcp-config <(echo '{"mcpServers":{}}') --strict-mcp-config
  

Common Pitfalls

• Missing Daemon: Forgetting to start the Ollama background daemon before running commands. Symptoms include connection refused on 127.0.0.1:11434.
• Process Substitution Syntax Errors: Using <(echo ...) requires bash or zsh. Running it in standard /bin/sh will result in a syntax error. Always ensure your execution shell is set to bash.
• Un-pulled Models: Attempting to launch an agent against a model that hasn't been cached locally will cause the agent initialization to fail or hang.
• Resource Exhaustion: Local models can consume significant RAM/VRAM. Monitor usage when running complex models like gemma4:e2b to avoid Out of Memory (OOM) kernel kills.

Troubleshooting

• Connection Refused: Start the Ollama server manually with ollama serve & if it's not running as a system service.
• Model Loading Hangs: If a model hangs during ollama run, check dmesg or system logs to ensure the system hasn't run out of memory.