Ollama

Metadata	Value
Category	ai
Capabilities	http, shell
Website	https://ollama.com

Returns shapes

• loaded_model[] — from ps
• model[] — from list_models, list_models_cli

Connections

• api — Ollama REST API — fast inference path, requires server to be running
• cli — Ollama CLI — can start the server, pull/delete models, management ops

Readme

Run local AI models — inference, tool calling, and model management — entirely on your machine. No API keys, no cloud, no cost per token.

Setup

Install Ollama (if not already installed):

brew install ollama
brew services start ollama

Pull a model to use:

ollama pull qwen3.5:9b-q8_0    # recommended: 11GB, tools + vision + thinking, 256K ctx
ollama pull glm-4.7-flash       # coding specialist: 19GB, best local SWE-bench

The skill auto-starts the Ollama server if it is not running — no manual setup required.

Connections

Connection	What it does
api (default)	REST API at localhost:11434 — fast inference, most operations
cli	Ollama CLI binary — server start/stop, model pulls, management

Usage

Tool	Connection	Description
status	cli	Check if server is running; start it if not
chat	api / cli	Multi-turn chat with tool calling and thinking mode
generate	api	One-shot text generation (faster for simple prompts)
list_models	api / cli	List all downloaded models with size and metadata
pull_model	cli / api	Download a model from the Ollama registry
delete_model	api / cli	Delete a model to free disk space
ps	api	Show models currently loaded in memory
show_model	api	Show model details: arch, context length, template

Tool calling

chat normalizes Ollama’s tool call format to the canonical AgentOS shape:

{ id, name, input }

Pass tools as { name, description, input_schema } — the same format used by the Anthropic skill.

Thinking mode

Models that support extended reasoning (qwen3, glm-4.7-flash, etc.) can be activated with thinking: true. The reasoning trace is returned in the thinking field, separate from content.

Notes

• chat with connection: cli collapses message history into a single prompt — suitable for single-turn only
• pull_model defaults to the CLI connection for reliable progress on large downloads (19GB models)
• ps shows what is currently warm in unified memory — useful before running a new model to estimate reload time
• The generate operation skips chat overhead — use it for classification, extraction, or quick single-turn tasks where speed matters