Ollama

Ollama is a local model runtime and control CLI for pulling models, running them locally or through Ollama Cloud, and exposing them over a local HTTP API. It also packages customized models and can launch supported coding tools against that runtime.

• Pull, run, stop, and inspect local or cloud-backed models from the shell, including interactive chat and embedding generation.
• Start a local Ollama server that exposes native and OpenAI-compatible JSON APIs for chat, embeddings, structured outputs, vision, and tool-calling requests.
• Create and import customized models from Modelfile, Safetensors, or GGUF assets, then launch supported tools like codex or claude against the local runtime.

• Agents usually get the most value from ollama serve plus the JSON API, with the CLI handling model lifecycle, setup, and simple one-shot runs.
• ollama run accepts piped stdin, supports --format json, and embedding models print a JSON array, so it can participate in shell pipelines even without wrapping the HTTP API.
• Fit is mixed rather than fully deterministic: the default entrypoint is a TUI, model responses are still probabilistic, and unattended use depends on local hardware limits or Ollama account auth for cloud flows.

• Large local models are constrained by available CPU, GPU, memory, and disk; cloud models require sign-in or API-key setup.
• The bare ollama and ollama launch flows are human-oriented Bubble Tea menus, so automation should call explicit subcommands or the API directly.