What is the Antigravity SDK?

The Antigravity SDK is Google's programmatic interface for embedding the same agent harness that powers Antigravity 2.0, Antigravity CLI, AI Studio, and Managed Agents in the Gemini API. Launched in preview at Google I/O 2026 on May 19, 2026, it lets third-party products integrate Google's agent runtime: planning, tool use, subagents, memory, checkpoints, and the manager-view orchestration model. Available initially for Python and TypeScript, optimized for Gemini models.

How is the Antigravity SDK different from the Gemini API?

The Gemini API is a model API — send tokens, get tokens back. The Antigravity SDK is an agent runtime — give it a task and tools, it plans, executes, calls models, manages state, and returns a result. You can think of the SDK as 'agent harness as a library' versus 'model as a service.' For multi-step work involving tool use, memory, and subagents, the SDK is the higher-level abstraction. For simple chat or single-call generation, the Gemini API is still the right tool.

When should I use Antigravity SDK vs Managed Agents in the Gemini API?

Use Managed Agents (the Interactions API endpoint) when you want Google to host the agent runtime — no infrastructure, pay-per-call. Use Antigravity SDK when you need to host the agent yourself: custom tools that can't be exposed over the wire, latency-sensitive workloads, on-prem deployments, or fine-grained control over the harness behavior. Same harness, two deployment models.

Is the Antigravity SDK open source?

No, the SDK is closed source — distributed as binary packages with documented APIs. Google has open-sourced parts of the surrounding tooling (some MCP servers, examples, the migration scripts from Gemini CLI) but the agent harness itself ships as a precompiled library. This contrasts with Anthropic's Claude Agent SDK (also closed) and OpenAI's Codex CLI (open). It's free to use in commercial products, subject to the Gemini API terms.

Quick Answer

What Is Antigravity SDK? Embed Google's Agent Harness (May 2026)

Published: May 21, 2026

What Is Antigravity SDK? Embed Google’s Agent Harness (May 2026)

The Antigravity SDK is the programmatic version of Antigravity — same planning, tool use, subagents, memory, and manager-view orchestration as the desktop app, available as a library you can drop into your own product. Launched in preview at Google I/O 2026 on May 19, 2026.

Last verified: May 21, 2026

Quick facts

Property	Value
Released	May 19, 2026 (preview at Google I/O)
Vendor	Google
Languages	Python, TypeScript (Node)
Models	Optimized for Gemini 3.5 Flash and Gemini 3.1 Pro
License	Closed source, free to use commercially
Pricing	Free SDK; pay-per-token Gemini API calls
Status	Preview — APIs may change before GA

What it actually is

The Antigravity SDK is the agent harness as a library. Same harness that powers:

Antigravity 2.0 (desktop app)
Antigravity CLI (terminal agent)
AI Studio (Google’s playground)
Managed Agents (Google-hosted agents in the Gemini API)

If you embed the SDK in your own product, you get the same agent runtime that powers Google’s own first-party products.

Why this exists

Google’s bet is that the agent runtime is the new platform. Before 2026, you’d glue together LangChain, your own tool registry, a planner you wrote, a memory layer, and a streaming UI. Each piece is fine; the integration is fragile.

The Antigravity SDK collapses that stack into a single library:

Old stack	Antigravity SDK
LangChain / LangGraph	Built in (planning + state)
Your tool registry	Built in (tools + MCP)
Your memory layer	Built in (session + persistent)
Streaming UI plumbing	Built in (content blocks)
Checkpoint / resume logic	Built in
Multi-agent orchestration	Built in (manager view)

Code shape

Python

from antigravity import Agent, Tool

agent = Agent(
    model="gemini-3.5-flash",
    instructions="You are a helpful research assistant.",
    tools=[Tool.from_mcp("github"), Tool.from_function(my_search_fn)],
)

result = await agent.run("Find issues in repo X labeled 'bug' from this week and summarize.")
print(result.output)

TypeScript

import { Agent, fromMCP, fromFunction } from "@google/antigravity";

const agent = new Agent({
  model: "gemini-3.5-flash",
  instructions: "You are a helpful research assistant.",
  tools: [fromMCP("github"), fromFunction(mySearchFn)],
});

const result = await agent.run("Find issues in repo X labeled 'bug' from this week and summarize.");
console.log(result.output);

Antigravity SDK vs alternatives

	Antigravity SDK	Claude Agent SDK	OpenAI Agents SDK	LangGraph
Vendor	Google	Anthropic	OpenAI	LangChain (OSS)
Open source	No	No	Yes	Yes
Default model	Gemini 3.5 Flash	Claude Opus 4.7	GPT-5.5	Bring your own
Multi-agent	Yes (manager view)	Yes (sub-agents)	Yes (handoffs)	Yes (graph)
Built-in MCP	Yes	Yes	Yes	Plugin
Built-in memory	Yes	Yes	Yes	Manual
Checkpoints	Yes	Yes	Yes	Yes (DeltaChannel)
Streaming UI primitives	Yes (content blocks)	Yes	Yes	Yes
Languages	Python, TS	Python, TS	Python, TS	Python (JS partial)
Tied to vendor model	Strong	Strong	Strong	None
Same harness as vendor products	Yes	Yes	Yes	n/a

The defining difference: Antigravity SDK gives you Google’s manager-view multi-agent model out of the box — agents that dispatch to subagents, run in parallel, and report to a coordinator.

When to use it

Use Antigravity SDK when…

You’re building an agent-powered feature in your own product
Your stack is Python or TypeScript
You want a single library instead of stitching together LangChain + a memory store + a tool registry
You’re already using Gemini models and want the harness optimized for them
You need multi-agent orchestration with parallel subagents
You want self-hosted deployment with full control over the runtime

Use Managed Agents (Gemini API) instead when…

You don’t want to manage agent runtime infrastructure
Your workload is bursty or low-volume — pay-per-call wins
You can expose your tools via MCP or simple HTTP — no need for in-process tools

Use Antigravity CLI / Antigravity 2.0 instead when…

The end user is you, a developer — not your product’s user
You want a GUI or TUI, not a library

What’s in preview

The SDK is preview as of May 19, 2026. That means:

APIs may change before GA
Some advanced harness features are not yet exposed (e.g., custom checkpoint backends)
Production SLAs aren’t promised yet
Pricing model for the harness itself (separate from API token costs) hasn’t been announced — currently free during preview

Google said GA is targeted for “later in 2026” — read that as Q3/Q4.

Pricing during preview

Component	Cost
SDK library	Free (preview)
Model calls (Gemini 3.5 Flash)	$1.50 / 1M input, $9 / 1M output
Model calls (Gemini 3.1 Pro)	Standard Gemini API rates
MCP servers (your own)	Whatever you pay to run them
Storage / state backend (your own)	Whatever you pay for Postgres / Redis

If you’re already on AI Pro or AI Ultra and developing personally, your subscription quotas count. For production workloads, expect to bill through a GCP project with Gemini API enabled.

Limits and caveats

Preview means breaking changes are possible — pin versions in production
Closed source — you can’t fork it. Compare with OpenAI’s Codex CLI (open) and LangGraph (MIT)
Tight coupling to Gemini — using Antigravity SDK with non-Google models works but isn’t optimized
Documentation is still thin — for May 2026, you’ll find more examples on the Antigravity GitHub repo than on the official docs

TL;DR

The Antigravity SDK is the agent harness behind every Google agent product, exposed as a Python/TypeScript library. If you’re building agent-powered features and want Google’s manager-view multi-agent model as a single dependency — instead of gluing together LangGraph + a memory store + a tool registry — the SDK is the option. Preview as of May 19, 2026; GA later this year. Free during preview; pay-per-token for the underlying Gemini calls.