Google Antigravity is Google's agent-first development platform — not AI as autocomplete in a sidebar, but autonomous agents with their own workspace that plan, execute, and verify software tasks across the editor, terminal, browser, and cloud, with human feedback at the decision points. Launched in public preview in November 2025 and expanded into a full ecosystem at Google I/O 2026, it now spans four surfaces — a desktop app, a CLI, a Python SDK, and the Gemini Enterprise Agent Platform — plus Managed Agents in the Gemini API. The headline for this tree: a first-class MCP surface. One caveat up front — it is not OSI open source: a proprietary compiled runtime, still a Research Preview.
Most AI coding tools bolt a model into a sidebar and call the human's turn after every suggestion. Antigravity inverts that. It gives agents their own workspace to plan and run end-to-end tasks — across the editor, terminal, and browser — and asks for human feedback only at the key decision points. The loop is idea → agent plans → executes across editor/terminal/browser → verifies → app, with the human in the loop at the checkpoints rather than on every keystroke.
The original product shipped (November 2025) as a fork of Visual Studio Code with two interaction modes: a synchronous Editor View — a familiar IDE with tab completion and inline commands — and an asynchronous Manager surface for orchestrating agents. A defining primitive is learning: agents save useful context and code snippets to a knowledge base so future tasks start smarter. At Google I/O 2026 (19 May 2026) Antigravity expanded from an IDE into a multi-surface ecosystem, all driven by the same underlying agent harness.
Antigravity is a Frameworks-band entry — an agent harness and platform, the natural Google sibling to Google ADK (the community has noted functional overlap between the two). It's also the destination of the Gemini CLI → Antigravity CLI migration covered in the coding-CLI leaf. This leaf is the platform view; that leaf is the CLI-surface view.
The 2026 expansion gives four ways to reach the same harness — pick the surface that fits the workflow, from a full desktop command center to a single API call.
The standalone command center — orchestrate multiple agents in parallel, spawn dynamic subagents, schedule cron-style background tasks, with integrations across AI Studio, Android, and Firebase.
Lightweight, keyboard-centric, high-velocity — creates agents instantly, no GUI. Google explicitly steers Gemini CLI users to migrate here.
A Python framework (Research Preview) with programmatic access to the same harness powering Google's own products — optimised for Gemini, hostable on your own infrastructure.
Connects Antigravity directly to Google Cloud projects, so enterprise workloads run inside an existing GCP estate.
Beyond the four surfaces, the Gemini API now exposes Managed Agents: a single call spins up an agent that reasons, uses tools, and executes code in an isolated Linux environment — powered by the Antigravity harness on Gemini 3.5 Flash. Gemini 3.5 Flash (launched at I/O 2026) is positioned as outperforming Gemini 3.1 Pro on almost all benchmarks while running roughly 4× faster — the speed engine for real-world agentic loops. Managed Agents accept custom instructions and skills written as markdown files.
For the 2nth.ai elastic-OS thesis this is the load-bearing finding. The Antigravity SDK exposes MCP servers as a first-class toolset — every SDK agent ships with built-in tools and can be extended with four tool types under one execution pipeline.
| Tool type | What it gives an agent |
|---|---|
| Built-in tools | File I/O, code editing, shell execution, directory search — out of the box |
| Custom Python functions | Register any Python callable as an agent tool |
| MCP servers | Connect any Model Context Protocol server — stdio, SSE, or HTTP |
| Agent Skills | Load reusable packages of instructions + tools, defined in markdown |
An agent can drive Antigravity SDK agents directly, and those agents can in turn drive any MCP server — a clean fit for the personal-agent → specialist-subagent nesting model. Tool execution is deny-by-default, governed by declarative policies that control when and how tools run. And because Agent Skills are markdown-defined, they line up with the same Agent Skills standard the 2nth-skills registry uses — wrap a specialist agent in well under ten lines and let it drive existing MCP surfaces.
The SDK quickstart is genuinely small — the API is deliberately agent-friendly (Pydantic V2 models, structured outputs, clean naming) so agents can read, write, and maintain SDK code as fluently as humans.
# pip install google-antigravity (pulls the compiled runtime binary) import asyncio from google.antigravity import Agent, LocalAgentConfig async def main(): config = LocalAgentConfig() async with Agent(config) as agent: response = await agent.chat("What files are in the current directory?") print(await response.text()) if __name__ == "__main__": asyncio.run(main())
The six lines are real, but the SDK depends on a compiled runtime binary shipped in platform-specific PyPI wheels. Cloning the public antigravity-sdk-python repo alone will not run it — you must pip install google-antigravity to fetch the binary. The quickstart's smallness is about API ergonomics, not about the SDK being a few hundred lines of open code.
Be precise about what Antigravity is and isn't. It is a powerful, MCP-native platform an agent can drive — it is not a self-hostable open-source dependency you can audit end to end.
antigravity-sdk-python repo is public, but the runtime is a compiled binary in PyPI wheels — the repo isn't standalone-runnable. The SDK is a Research Preview, an evolving platform, not a stability guarantee.Treat Antigravity as a proprietary-runtime, agent-driveable platform — excellent for build velocity and as an MCP client/host, but not a dependency you can self-host on open weights today. If a self-hostable, auditable runtime is the requirement, reach for ADK (or another open framework) instead, and revisit Antigravity when the Gemma-on-open-models roadmap item ships.
The proprietary compiled runtime and Research-Preview status matter for SA procurement and POPIA reviews: Antigravity is a velocity tool and MCP host, not a self-hostable dependency you can audit end to end. Managed Agents run in Google-managed Linux environments — the in-country, open-weights path (the Gemma runtime) is roadmap, not shipped. For clients with data-residency constraints, the Gemini Enterprise Agent Platform connection is the route to keep workloads inside an existing Google Cloud project (and, where available, an SA region).
Strong fit for SA studios chasing delivery velocity: the SDK's MCP-server toolset and markdown Agent Skills line up with the 2nth-skills registry pattern (npx skills add …). A studio can wrap a specialist agent in under ten lines and let it drive existing MCP surfaces — the personal-agent → subagent-team nesting model, in practice. A good candidate for an agents/ building block, flagged as proprietary-runtime.