v0.2.1

Four PRs on top of v0.2.0, all about local and provider setup: Qwen 3.8 27B in the managed catalog within days of its weights opening, a first-class Gemini provider, a fix for Qwen models that emit tagged tool calls, and a local Ollama preset. The Gemini provider and the Qwen tool-call fix are community contributions from @yablokolabs, an external contributor.

Highlights

Qwen 3.8 27B, day-one in the catalog

Qwen 3.8 is Alibaba's newest open-weight release, and its 27B local build is in the managed catalog here within days of the weights going public. Add it from the model list with no manual setup, it runs on consumer hardware (roughly 17 GB of RAM or VRAM), and it brings the Qwen and Gemma count in the managed catalog to 10. If you want the latest open model running locally under Atomic Agent, it is one pick away. (#127)

First-class Gemini provider

Until now, running Gemini meant setting it up by hand as a generic OpenAI-compatible endpoint. That path did not actually work without a translating proxy, because Atomic Agent appends /v1/... to the base URL while Google's OpenAI-compatible API lives under /v1beta/openai/.... This release adds Gemini as a proper provider in the setup wizard. Pick it from the list and you get a stable default chat model with no manual base-URL step, chat completions and model discovery are routed through Google's documented /v1beta/openai/... paths, and your key is read from GEMINI_API_KEY (never written to logs or errors). The existing OpenAI-compatible, OpenRouter, and AIMLAPI paths are untouched. This one comes from @yablokolabs, thank you for the contribution. (#119, closes #108)

A smoother local setup

Two more changes make local models easier to run.

• A local Ollama preset. Pick ollama in the provider wizard for a server running under ollama serve at http://localhost:11434. It has no API key, so the wizard skips both the base-URL screen and the key screen: you go straight to picking a model. This is separate from Ollama Cloud, which points at the hosted service. (#122)
• A fix for Qwen builds that emit tagged tool calls (see below), so more Qwen deployments drive the agent loop without stalling. (#120)

Everything else

Providers and models

• First-class Gemini provider in the setup wizard, routed through Google's /v1beta/openai/... paths, with credentials read from GEMINI_API_KEY. Existing provider paths are unchanged. (#119, closes #108)
• Local Ollama preset (ollama, http://localhost:11434, no key), separate from the hosted Ollama Cloud preset. Picking it skips the base-URL and key screens. (#122)
• Qwen 3.8 27B added to the managed local model catalog. (#127)

Fixed

• Qwen tagged tool calls. Some Qwen deployments behind OpenAI-compatible endpoints return tool calls as tagged text inside content or reasoning_content instead of native message.tool_calls, and Atomic Agent used to treat those valid calls as plain assistant prose, which stalled the agent loop. An opt-in, provider-specific Qwen adapter now parses the tagged calls when native ones are absent, resolves dotted tool names (like os.fs.read) against the tools actually offered, and validates the parsed parameters against the offered schema. It fails closed on malformed tags or unknown tools, and ordinary prose stays prose. Generic OpenAI, OpenRouter, AIMLAPI, and Gemini behavior is not affected. This is a community contribution from @yablokolabs. (#120, closes #105)

Install / Update

macOS / Linux:

curl -fsSL https://atomicagent.io/install | sh

Windows (PowerShell):

irm https://atomicagent.io/install.ps1 | iex

Atomic Agent updates itself in place. After an update the TUI prompts you to restart.

Full Changelog: v0.2.0...v0.2.1

Gemini support and the Qwen tool-call fix come from @yablokolabs, thanks for the contribution. Built by @yablokolabs, @sosidudku1, and @Ooooze.