Ollama has changed what happens when you run the bare ollama command with no arguments. Two releases five days apart move the project from a simple local chat prompt toward something closer to a coding agent that runs entirely on your own hardware: v0.32.0, published July 11, 2026, and v0.32.1, published five days later on July 16, 2026.
From chat to agent by default
Before v0.32.0, running ollama with no arguments dropped you into a plain chat prompt. That default now launches an interactive agent instead, built to write and edit code and delegate tasks rather than just answer questions turn by turn. Anyone running Ollama as a local API server behind Open WebUI or another front end won’t notice a difference there. But the bare CLI itself now behaves less like a chat prompt and more like an assistant that expects to operate inside a project directory.
Codex becomes ChatGPT, and the launch menu shrinks
The same release renames the “Codex App” integration to “ChatGPT.” It’s reached with ollama launch chatgpt, and a new --restore flag switches back to your usual ChatGPT profile afterward. The ollama launch menu has also been trimmed: only the most popular integrations show up by default, with the rest still reachable by running the command and browsing further instead of being listed up front. It’s a small change, but it suggests Ollama wants launch to work as a curated starting point rather than a full catalog every time.
Deprecation warnings for older agent models
v0.32.0 also adds a deprecation warning that fires before you launch older models in agent mode. It covers:
- CodeLlama
- Qwen2.5 and Qwen2.5-coder
- The Llama 3.x line
- Mistral
- StarCoder
- The base DeepSeek-R1 tag
None of these models are removed. But the CLI now tells you upfront that they’re on their way out as agent-capable defaults, worth knowing if you have automation or scripts pinned to one of them.
Five days later, v0.32.1 tightens the rough edges
A release that changes the default interaction model this much tends to get a fast follow-up. v0.32.1 reads like exactly that: a set of fixes aimed at the friction points the new agent flow surfaced almost immediately.
The most user-facing fix improves Gemma 4’s tool calling and multi-turn reasoning, making tool-call continuations more reliable when a task takes several turns to resolve instead of one shot. That matters for the new agent mode specifically, since delegating work to a model only works as well as the model’s ability to keep calling tools correctly partway through a job.
The rest of the release is a cleanup pass:
- Fixed a recurring memory leak in the MLX model cache and improved cache snapshot performance, relevant to Apple Silicon users
- MLX text model loading now respects the
OLLAMA_LOAD_TIMEOUTsetting, so slow-loading models aren’t killed prematurely on constrained hardware - The agent’s web search and fetch tool now prompts
ollama signinwhen authentication is actually required, instead of failing without explanation - The interactive agent now receives the current working directory, giving it better context about the project it’s running in
- Fixed the model picker in
ollama launch: the “Pick another model” option now works correctly when a deprecated model is passed via--model - Updated documentation for configuring the official Ollama extension in VS Code
Why this matters for self-hosted setups
Ollama’s appeal for homelabbers has always been running capable models on your own hardware without sending prompts to a third-party API. Turning the bare CLI into an agent by default is a bigger shift than a typical point release. It changes what people get when they type ollama into a terminal, and it pushes the tool further into coding-assistant territory, alongside tools like Aider or IDE copilots, except running fully local.
If you rely on Ollama purely as an API backend for Open WebUI or a similar stack, nothing in that setup breaks. This change affects the bare CLI’s default behavior, not the API server most self-hosted front ends actually talk to. But if you run the ollama command directly on the box, the experience going forward looks different from what you’re used to.
Updating
Both releases are already out. Updating through the official install script or your package manager brings you to v0.32.1, which includes everything from v0.32.0 plus the July 16 fixes. Docker users running the ollama/ollama image should re-pull latest or pin to 0.32.1 explicitly. Since this release changes default CLI behavior, it’s worth checking any scripts or automation that call ollama directly before rolling the update out broadly.