October agent roundup: OpenAI, Google and GitHub all shipped CLI agent updates.

All three big labs, OpenAI, Google and GitHub, released agent updates in October, and the open-source agents did too. I wanted one place to compare the practical side: local models, accounts, cost, and what each one sends home by default. Agent What shipped in October Runs on a local model Needs a cloud account License Price Usage stats by default Codex CLI (OpenAI) GPT-6.1 Terra (Oct 26), CLI 0.160 to 0.166 partly: --oss with Ollama or LM Studio, or a custom endpoint yes for OpenAI models CLI Apache-2.0 included in ChatGPT Free and Go ($8/mo), more usage on Plus ($20/mo) and up, or API key on, can turn off Gemini CLI (Google) CLI 0.63 to 0.66 (Oct 6 to 27) no: only an optional, experimental local routing model yes: Code Assist Standard/Enterprise login or a paid API key CLI Apache-2.0 no free tier since June 18 (free users moved to Antigravity CLI) on, can turn off Copilot CLI (GitHub) local model discovery (Oct 7), Copilot Fusion (Oct 28) yes: Ollama or any OpenAI-compatible endpoint not with your own key, otherwise yes closed source free tier, Pro $10/mo on even with your own key, can turn off Hermes Agent (Nous) v0.22 "Threshold" (Oct 13): one-click local models yes, built-in llama.cpp no MIT free off (opt-in) Atomic Agent v0.6.7 to 0.6.13, desktop app (Oct 7 to 28) yes, built-in llama.cpp by default no MIT free on (anonymous), can turn off OpenClaw v2026.10.1 (Oct 10) to 2026.10.9 yes: llama.cpp plugin, Ollama, LM Studio no MIT free daily update check on, feature stats opt-in What stood out The big labs went local this month, but only halfway. Copilot CLI now finds your Ollama models on its own, and Codex has --oss. Gemini CLI is the odd one out: the agent itself always talks to Google, and only an optional, experimental routing model can run locally. "Local" doesn't mean quiet. Copilot CLI's docs say it still sends telemetry even when you bring your own key, unless you switch on offline mode. Codex, Gemini CLI and Atomic Agent also send usage stats by default. Hermes is the only one here where it's off until you turn it on. The three open-source ones (Hermes, OpenClaw, Atomic Agent) all run without any account, all ship their own llama.cpp setup now, and all are MIT. One surprise: Gemini CLI's terms list using OpenClaw with Gemini CLI's login as a violation that can get your account suspended. Prices and features are from each project's own docs and changelogs as of Oct 30. Which one are you using, and is fully offline a must-have or a nice-to-have for you? Where this comes from All public, from each project's own GitHub repo, docs and changelog: Codex CLI: openai/codex on GitHub, plus OpenAI's Codex changelog, config and pricing docs Gemini CLI: google-gemini/gemini-cli on GitHub, plus the Gemini CLI terms and privacy page Copilot CLI: github/copilot-cli on GitHub, plus GitHub's Copilot CLI docs and the Oct 7 and Oct 28 changelog posts Hermes Agent: NousResearch/hermes-agent on GitHub (v0.22.0 release notes) Atomic Agent: AtomicBot-ai/atomic-agent on GitHub OpenClaw: openclaw/openclaw on GitHub, plus the OpenClaw providers docs