# Antigravity CLI and OpenCode 2 > NeuroSquad 0.1.283 adds Antigravity CLI, Google’s successor to Gemini CLI, with the full integration, moves Gemini CLI cards over in one click, and supports OpenCode 2 next to 1.x with dangerous mode that switches live. - URL: https://neurosquad.ai/en/blog/antigravity-cli-and-opencode-2/ - Published: 2026-10-08 - Publisher: NeuroSquad (https://neurosquad.ai/) NeuroSquad 0.1.283 brings two changes to the agent CLIs it runs. **Antigravity CLI** — Google’s `agy`, the successor to Gemini CLI — joins with the full integration and takes Gemini CLI’s place in the add-card menu, the setup wizard and templates. And **OpenCode 2**, a new implementation behind the same `opencode` command, now works in cards next to OpenCode 1.x. The app still runs [22 agent CLIs](https://neurosquad.ai/en/harnesses/). This is a Windows release; the Mac app stays on 0.1.230 for now and gets both with its next build. ## Antigravity CLI, in Gemini CLI’s place Google closed Gemini CLI to individual accounts in June and points its users to Antigravity CLI. For NeuroSquad that meant a new integration, not a rename: `agy` is a single closed-source binary, so everything the card does was learned from Google’s documentation, the CLI’s changelog, its own help and live probes. As with every CLI before it, the work went row by row through the checklist of what the app does for Claude Code — Antigravity’s document has 45 rows, each marked native, adapted or impossible. The core of it is one mechanism. Antigravity keeps everything under `~/.gemini` and offers no variable to move its hooks or MCP settings, but the binary accepts a directory of its own at launch. So every card runs with **its own Antigravity folder**: settings made of yours plus the card’s layer, hooks and MCP servers made of yours plus NeuroSquad’s, and everything else — conversations, transcripts, plugins, skills — linked to your real folder, which is never written. A conversation started in a card can be continued from a plain terminal and back. On top of that comes what makes it a full card. **Statuses** come from Antigravity’s own hooks and the state it reports for its window title: working, needs input with the command or question it waits on, finished — and a turn stays open while its subagents are still reporting. **Conversations resume** after a restart. The app’s MCP tools arrive **by arrow** and are re-listed live; connected MCP servers are asked for on every call. Escape interrupts (Ctrl+C is its exit key). **Usage**, Run Stats and Agent Pulse read its transcript, subagents included, and the card shows a context chip and a quota chip. The model and CLI switch, templates, and plugin context for Caveman, Memory, Context7 and Graphify all work through the same hooks. ![An Antigravity CLI card marked Working: the agy banner with version 1.3.1, the prompt “Read README.md and suggest three quick improvements” answered with three numbered suggestions, then a second prompt “Now write the Quick start section into README.md” with one file explored and “Reading file…”](https://neurosquad.ai/blog/antigravity-cli-and-opencode-2/antigravity-en.webp) *An Antigravity CLI card mid-turn. The model here is a scripted test server.* > **Found on the way: an empty answer that says no** Antigravity’s documentation describes hooks in its settings file; the CLI only reads them from a separate hooks file, which is where the card puts them. And a pre-tool hook that answers with an empty JSON object **refuses** the tool rather than staying neutral — so the app answers with nothing at all unless canvas mode has to refuse. Some rows stay empty, and the document says why. Antigravity reads neither its settings nor a hook’s approval while it runs, so **dangerous mode applies when the card restarts** (with the same conversation). There is no Compact button and no RTK, and Antigravity cards don’t run in WSL or SSH workspaces yet. Install it from the setup wizard. ## Gemini CLI cards: one press to move Gemini CLI is no longer offered for new cards, but nothing you already have stops working: existing Gemini CLI cards launch exactly as before, now labelled **Gemini CLI (legacy)**. Each shows a small note with **Switch to Antigravity** — one press runs the harness switch in place and carries the conversation’s context into the new Antigravity session; if a turn is running, the switch waits for it to end. Templates that contain Gemini CLI cards import them as Antigravity CLI cards, and the import dialog says so. ![A Gemini CLI card with the note “Gemini CLI is retired — Antigravity CLI replaced it” and a blue “Switch to Antigravity” button above the running Gemini CLI prompt](https://neurosquad.ai/blog/antigravity-cli-and-opencode-2/legacy-en.webp) *A Gemini CLI card keeps running and offers the move.* ## OpenCode 2, next to 1.x OpenCode 2 is a separate implementation behind the same command, with a new plugin API, a new config schema and its own session tables. Its official instructions still install 1.x, so both will be on people’s machines for a while. A card now asks the binary it is about to start for its version and picks the launch to match — **1.x runs exactly as before**, and a 2.x you installed yourself is used as it is. On 2.x several things got better than on 1.x. **Dangerous mode switches live**: the NeuroSquad plugin answers OpenCode’s own permission check, so turning it on or off applies to the next question without a restart. **The app picks the session id** before the first turn — OpenCode 2 creates a session under an id it is given — so the card knows its session from the start and never has to recover from a missing one. A model change made by you or a lead agent applies to the same session. The app’s tools **appear live** as you draw arrows. And requests to OpenRouter carry NeuroSquad’s attribution **also when you use your own OpenRouter login**, which OpenCode would otherwise send as its own. ![An OpenCode 2 card: the session tab “README improvements”, the prompt “Read README.md and suggest three quick improvements”, one file read, three numbered suggestions, and the Build agent line with the model and timing](https://neurosquad.ai/blog/antigravity-cli-and-opencode-2/opencode-en.webp) *An OpenCode 2 card after one turn, on a local test model.* > **Found on the way: tools behind one door, a plugin that restarts** By default OpenCode 2 shows an MCP server’s tools to the model only through a single “execute” tool, so a card’s config turns that off for the app’s servers — otherwise the tools you give by arrow would be hidden. And OpenCode 2 watches its plugin files: rewriting the shared plugin folder reloaded the plugin in every running 2.x card, mid-turn. The folder is now written only when its content actually changes. ## How it was tested Both CLIs were installed into scratch folders, never globally, and driven by a separate copy of the app with its own profile, a scratch home folder and scripted model servers on the loopback interface — no one’s real login or config was read. For Antigravity that meant its API-key mode pointed at a scripted Gemini-API server, and a second card going through the app’s gateway to a custom provider; the run passed all 51 checks, and the files in the scratch `~/.gemini` were byte-identical afterwards. OpenCode was run twice through the same 60 checks, once on 2.0.24 and once on 1.18.32. Usage was compared with the model servers’ own logs, request by request. The screenshots in this post come from the same kind of setup. Each CLI has its own page: [Antigravity CLI](https://neurosquad.ai/en/harnesses/antigravity/) and [OpenCode](https://neurosquad.ai/en/harnesses/opencode/), with setup details in the [docs](https://docs.neurosquad.ai/en/agents/antigravity). The full list of changes is in the [0.1.283 changelog](https://neurosquad.ai/en/changelog/#v0.1.283). NeuroSquad updates itself; a new install starts from the [download page](https://neurosquad.ai/en/download/).