Ollama Is an Agent Now: v0.32 Turns the Bare Command Into Chat, Code, and Web Search

Type ollama with no arguments after updating to v0.32 and you no longer get a help page: you get an interactive agent that chats, writes code, searches the web, and, in Ollama's words, delegates real work. The tool most of us installed to run models locally is now an agent harness. Three releases shipped in ten days, and one default deserves your attention before you press enter.

What's new in v0.32

• The bare command is an agent (v0.32.0, July 11): running ollama launches "Chat, Code, & Work", a session that can chat with models, edit code, search the web, and run multi-step tasks. In v0.32.1 it picks up your current working directory, so it operates on the project you are standing in.
• Skills and longer tool runs (v0.32.2, July 20): the agent gained a skills system and unlimited tool rounds for cloud models, plus a Claude Code launch integration alongside the existing ChatGPT one. On Linux, CUDA compute capability 10.0 support arrived.
• Housekeeping that matters (v0.32.1, July 16): an MLX engine cache leak that grew memory across requests on Apple Silicon is fixed, and older agent models (CodeLlama, Qwen2.5-coder, Llama 3.x, Mistral, StarCoder, base DeepSeek-R1) now carry deprecation warnings.

Why it matters for open-source builders

Every terminal agent so far has asked you to install something new. Ollama just gave one to everybody who already has it installed, which is a very large number of machines. But read the model name in the banner: the default agent runs on glm-5.2:cloud, Ollama's hosted deployment of GLM-5.2, and web search requires an ollama signin. That means the out-of-the-box experience of the local-AI tool sends your prompts, and potentially your code, through Ollama's cloud. The launcher has a model picker, so you can point the agent at a local model instead; just know that the default is no longer the thing Ollama's name stands for. We flagged the same drift when the web search API landed in v0.31. The pattern is now unmistakable.

Try it in two minutes

# Update, then launch the agent
curl -fsSL https://ollama.com/install.sh | sh   # macOS/Windows: update the app
ollama

# Prefer a local model? Pick one in the launcher,
# or keep using the classic commands, unchanged:
ollama run gemma3

One-click links: release notes on GitHub · download Ollama · docs

Update, type ollama, and ask it to summarize the repo you are sitting in. Then open the model picker and see how far a local model gets on the same task. That comparison, on your own hardware, is the most honest benchmark you will read this week.

All changes are as documented in Ollama's official release notes for v0.32.0 through v0.32.2; we have not yet load-tested the agent mode ourselves.
Date checked: 2026-07-22