Building on the basics of an AI Harness: Antigravity CLI is the terminal interface of Google Antigravity, the company’s agent-first development platform for coding agents with selectable Gemini, Claude and GPT-OSS models. As of 20 September 2026.
What is Antigravity CLI?
Per its own documentation, Google Antigravity is a development platform made up of several surfaces: the “Antigravity 2.0” desktop application as a command centre for multiple agents, the IDE, a Python SDK for building custom agents, and Antigravity CLI itself as a “lightweight, keyboard-centric terminal interface”. Per the documentation, Antigravity CLI brings “the reasoning, execution, and orchestration capabilities of our shared agent harness” directly to the local shell — the same harness that Antigravity 2.0 runs on. Antigravity CLI is installed on macOS/Linux via curl -fsSL https://antigravity.google/cli/install.sh | bash, on Windows via a PowerShell script; the session starts with the agy command. A public GitHub repository (google-antigravity/antigravity-cli) contains installation scripts and points to the documentation but lists no open-source licence — Antigravity CLI is a proprietary Google product. Per Google’s official announcement of 19 May 2026, Antigravity CLI is built in Go, intended to enable faster execution than Gemini CLI.
Antigravity CLI as an AI harness
Antigravity CLI’s agent loop is shared with the Antigravity 2.0 desktop application: improvements to reasoning or code understanding, per the documentation, apply automatically to both surfaces. Tools are managed via an interactive MCP manager, opened with /mcp, and a terminal sandbox mechanism using native OS features (Linux namespaces, sandbox-exec/Seatbelt profiles on macOS) guards against destructive file operations and unauthorised network access — on Windows, the documentation so far only describes access governed by approval rules, without a dedicated OS-level isolation technology. Project-level instructions live as Markdown rules in the repository’s .agents/rules/ directory, with global instructions additionally in ~/.gemini/GEMINI.md or ~/.gemini/antigravity-cli/rules/; agent skills encapsulate recurring workflows as slash commands. An asynchronous subagent framework delegates parallel work, background research and system tests without blocking the active conversation — monitored and controlled via the /agents command. Plugins bundle skills, agents, rules, MCP servers and hooks into a single package, auto-loaded from ~/.gemini/antigravity-cli/plugins/. Sessions can be resumed via /resume, rolled back to an earlier checkpoint via /rewind (alias /undo), and exported to the Antigravity 2.0 desktop interface.
Models and model access
Per the documentation, Antigravity CLI offers a model selector covering models from the Gemini 3 family as well as — depending on the plan — Claude models and GPT-OSS; an additional, non-customisable image model, “Nano Banana 2”, is available for UI mockups and diagrams. Enterprise licences, per the documentation, are explicitly limited to Gemini models, with Claude and GPT-OSS unavailable there. A freely configurable, provider-agnostic endpoint is not documented in Antigravity CLI’s configuration reference (settings.json); whether model access can be routed through a central proxy such as the AI Gateway is therefore not publicly documented.
Enterprise use: data protection and governance
For enterprise licences, Google documents that prompts, responses, code and telemetry are never stored outside your own private Google Cloud environment, and that customer telemetry is logged directly to the project of the chosen licence. Besides a global endpoint, the documentation explicitly names an EU multi-region for text generation, code inference and multimodal requests — image generation, by contrast, runs exclusively via global endpoints. Sign-in happens via Google Cloud business accounts through single sign-on, complemented by Workforce Identity Federation for external identity providers such as Okta. VPC Service Controls enable private network paths, and request/response logs are available for audits. Outside enterprise use, the GitHub repository states that interaction data is sent to Google by default to improve the product, with an opt-out available in settings. Regardless of the access route, source code and prompts go to the chosen model provider and belong in the record of processing activities beforehand.
Orchestrating Antigravity CLI with herdr
The terminal multiplexer herdr lists Antigravity CLI in its table of supported agents with screen manifest as the state authority and session-level integration — the same pattern as Claude Code, Codex or Cursor Agent CLI. This lets Antigravity CLI run alongside other coding agents inside a herdr-managed terminal session.
Where it fits: developer harness and enterprise harness
Like other coding harnesses, Antigravity CLI is a tool for developers working in the terminal or alongside the desktop IDE. For business users without a terminal workflow, CompanyGPT is the matching harness — with tools via MCP servers, integrations and n8n. The AI Gateway provides the shared cost and governance layer across all model access in an organisation. An overview of further harnesses is available on the AI harness overview page. We help organisations fit Antigravity CLI and other coding harnesses into an existing AI governance concept.