Antigravity Agent

The Antigravity agent is a general-purpose managed agent on the Gemini API. A single API call gives you an agent that reasons, executes code, manages files, and browses the web inside your own secure Linux sandbox, hosted by Google.

It is powered by Gemini 3.5 Flash and uses the same harness as the Antigravity IDE. Available through the Interactions API and Google AI Studio.

Python

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="Read Hacker News, summarize the top 10 stories, and save the results as a PDF.",
    environment="remote",
)

print(interaction.output_text)

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "Read Hacker News, summarize the top 10 stories, and save the results as a PDF.",
    environment: "remote",
}, { timeout: 300000 });

console.log(interaction.output_text);

REST

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "agent": "antigravity-preview-05-2026",
    "input": "Read Hacker News, summarize the top 10 stories, and save the results as a PDF.",
    "environment": "remote"
}'

Capabilities

Each call can provision a Linux sandbox and starts a tool-use loop. The agent plans, acts, observes results, and repeats until the task is done.

  • Code execution: Run Bash, Python, and Node.js commands. Install packages, run tests, build apps.
  • File management: Read, write, edit, search, and list files in the sandbox. Files persist across interactions.
  • Web access: Google Search and URL fetching for data.
  • Context compaction: Automatic context compaction (triggered at ~135k tokens) to support long-running, multi-turn sessions without losing context or hitting token limits.

See the Quickstart for multi-turn usage and streaming.

Supported tools

By default, the agent has access to code_execution, google_search, and url_context. Filesystem tools are enabled automatically when you specify the environment parameter. You can also define custom functions to connect the agent to your own APIs and tools. You only need to specify the tools parameter when customizing or restricting the default set, or when adding custom functions.

Tool Type value Description
Code Execution code_execution Run shell commands (bash, Python, Node) with stdout/stderr capture.
Google Search google_search Search the public web.
URL Context url_context Fetch and read web pages.
Filesystem (enabled via environment) Read, write, edit, search, and list files in the sandbox. No separate tool type; enabled automatically when environment is set.
Custom Functions function Define custom functions that the agent can request to execute. See Function calling.
Remote MCP Server mcp_server Register external Model Context Protocol (MCP) servers as tools. See MCP servers.

To limit the agent to specific tools, pass only the ones you need:

Python

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="Search for the latest AI research papers on reasoning and summarize them.",
    environment="remote",
    tools=[
        {"type": "google_search"},
        {"type": "url_context"},
    ],
)

print(interaction.output_text)

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "Search for the latest AI research papers on reasoning and summarize them.",
    environment: "remote",
    tools: [
        { type: "google_search" },
        { type: "url_context" },
    ],
}, { timeout: 300000 });

console.log(interaction.output_text);

REST

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "agent": "antigravity-preview-05-2026",
    "input": "Search for the latest AI research papers on reasoning and summarize them.",
    "environment": "remote",
    "tools": [
        {"type": "google_search"},
        {"type": "url_context"}
    ]
}'

Multimodal Input

The Antigravity agent supports multimodal inputs. Currently, only text and image inputs are supported. Images must be supplied as inline base64-encoded strings (data).

Python

import base64
from google import genai

client = genai.Client()

with open("path/to/chart.png", "rb") as f:
    image_bytes = f.read()

interaction_inline = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input=[
        {"type": "text", "text": "Analyze this chart and summarize the trends."},
        {
            "type": "image",
            "data": base64.b64encode(image_bytes).decode("utf-8"),
            "mime_type": "image/png",
        },
    ],
    environment="remote",
)

JavaScript


import { GoogleGenAI } from "@google/genai";

import * as fs from "node:fs";

const client = new GoogleGenAI({});
const base64Image = fs.readFileSync("path/to/chart.png", { encoding: "base64" });

const interactionInline = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: [
        { type: "text", text: "Analyze this chart and summarize the trends." },
        {
            type: "image",
            data: base64Image,
            mime_type: "image/png",
        },
    ],
    environment: "remote",
}, { timeout: 300000 });

REST

BASE64_IMAGE=$(base64 -w0 /path/to/chart.png)

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d "{
    \"agent\": \"antigravity-preview-05-2026\",
    \"input\": [
        {\"type\": \"text\", \"text\": \"Analyze this chart and summarize the trends.\"},
        {
            \"type\": \"image\",
            \"mime_type\": \"image/png\",
            \"data\": \"$BASE64_IMAGE\"
        }
    ],
    \"environment\": \"remote\"
}"

Function calling

Function calling allows you to connect the Antigravity agent to external APIs and databases by defining custom tools the agent can invoke. For general concepts, see Function calling with the Gemini API.

The following example demonstrates a 2-turn interaction. The agent first requests a custom get_weather function call, and the client executes it and returns the result in the second turn.

Python

from google import genai

client = genai.Client()

# 1. Define the custom function
get_weather_tool = {
    "type": "function",
    "name": "get_weather",
    "description": "Gets the current weather for a given location.",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "The city and country, e.g. San Francisco, USA",
            }
        },
        "required": ["location"],
    },
}

# 2. Call the agent with the custom tool (Turn 1)
interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="What is the weather in Tokyo?",
    environment="remote",
    tools=[
        {"type": "code_execution"},  # Enable default code execution
        get_weather_tool,            # Add custom function
    ],
)

# Check if the agent requested a function call
if interaction.status == "requires_action":
    # Find function calls that do not have a matching function result.
    # Filesystem tools (like write_file) are also represented as function calls
    # but are executed automatically by the environment.
    executed_calls = {step.call_id for step in interaction.steps if step.type == "function_result"}
    pending_calls = [step for step in interaction.steps if step.type == "function_call" and step.id not in executed_calls]

    if pending_calls:
        fc_step = pending_calls[0]
        print(f"Function to call: {fc_step.name} (ID: {fc_step.id})")
        print(f"Arguments: {fc_step.arguments}")

        # 3. Execute the function locally (simulated get_weather()) and send the result back (Turn 2)
        function_result = {
            "temperature": 23,
            "unit": "celsius"
        }

        final_interaction = client.interactions.create(
            agent="antigravity-preview-05-2026",
            previous_interaction_id=interaction.id,  # Reference the interaction ID
            environment=interaction.environment_id,
            input=[
                {
                    "type": "function_result",
                    "name": fc_step.name,
                    "call_id": fc_step.id,
                    "result": function_result,
                }
            ],
        )

        print(final_interaction.output_text)
        # Output: The current weather in Tokyo, Japan is 23°C (Celsius).
    else:
        print("No pending function calls.")
else:
    print(f"Interaction completed with status: {interaction.status}")

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

// 1. Define the custom function
const get_weather_tool = {
  type: "function",
  name: "get_weather",
  description: "Gets the current weather for a given location.",
  parameters: {
    type: "object",
    properties: {
      location: {
        type: "string",
        description: "The city and country, e.g. San Francisco, USA",
      },
    },
    required: ["location"],
  },
};

// 2. Call the agent with the custom tool (Turn 1)
const interaction = await client.interactions.create({
  agent: "antigravity-preview-05-2026",
  input: "What is the weather in Tokyo?",
  environment: "remote",
  tools: [
    { type: "code_execution" },
    get_weather_tool,
  ],
}, { timeout: 300000 });

if (interaction.status === "requires_action") {
  // Find function calls that do not have a matching function result.
  // Filesystem tools (like write_file) are also represented as function calls
  // but are executed automatically by the environment.
  const executedCalls = new Set(
    interaction.steps
      .filter(s => s.type === "function_result")
      .map(s => s.call_id)
  );
  const pendingCalls = interaction.steps.filter(
    s => s.type === "function_call" && !executedCalls.has(s.id)
  );

  if (pendingCalls.length > 0) {
    const fcStep = pendingCalls[0];
    console.log(`Function to call: ${fcStep.name} (ID: ${fcStep.id})`);

    // 3. Execute the function locally (simulated get_weather()) and send the result back (Turn 2)
    const functionResult = {
      temperature: 23,
      unit: "celsius"
    };

    const finalInteraction = await client.interactions.create({
      agent: "antigravity-preview-05-2026",
      previous_interaction_id: interaction.id, // Reference the interaction ID
      environment: interaction.environment_id,
      input: [
        {
          type: "function_result",
          name: fcStep.name,
          call_id: fcStep.id,
          result: functionResult,
        }
      ],
    }, { timeout: 300000 });

    console.log(finalInteraction.output_text);
  } else {
    console.log("No pending function calls.");
  }
} else {
  console.log(`Interaction completed with status: ${interaction.status}`);
}

REST

# 1. Turn 1: Request function call
RESPONSE=$(curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -d '{
      "agent": "antigravity-preview-05-2026",
      "input": "What is the weather in Tokyo?",
      "environment": "remote",
      "tools": [
          {"type": "code_execution"},
          {
              "type": "function",
              "name": "get_weather",
              "description": "Gets the current weather for a given location.",
              "parameters": {
                  "type": "object",
                  "properties": {
                      "location": {"type": "string"}
                  },
                  "required": ["location"]
              }
          }
      ]
  }')

# Extract interaction ID, environment ID, and call ID (requires jq)
INTERACTION_ID=$(echo $RESPONSE | jq -r '.id')
ENVIRONMENT_ID=$(echo $RESPONSE | jq -r '.environment_id')
CALL_ID=$(echo $RESPONSE | jq -r '.steps[] | select(.type=="function_call") | .id')

# 2. Turn 2: Send function result back using variables
curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -d "{
      \"agent\": \"antigravity-preview-05-2026\",
      \"previous_interaction_id\": \"$INTERACTION_ID\",
      \"environment\": \"$ENVIRONMENT_ID\",
      \"input\": [
          {
              \"type\": \"function_result\",
              \"name\": \"get_weather\",
              \"call_id\": \"$CALL_ID\",
              \"result\": {
                  \"temperature\": 23,
                  \"unit\": \"celsius\"
              }
          }
      ]
  }"

MCP servers

You can connect the Antigravity agent to external tools by registering remote Model Context Protocol (MCP) servers. The agent supports remote MCP servers over streamable HTTP.

When registering an MCP server, you must specify the following fields in the tools array:

Field Type Required Description
type string Yes Must be "mcp_server".
name string Yes A unique identifier for the server. Must be strictly lowercase and alphanumeric (matching ^[a-z0-9_-]+$).
url string Yes The endpoint URL of the remote MCP server.
headers object No Custom headers (e.g., authentication) sent with requests.
allowed_tools array No List of tool names allowed to be executed. If omitted, all tools are allowed.

Python

from google import genai

client = genai.Client()

# Register a remote HTTP MCP server
interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="What is the weather in Tokyo?",
    environment="remote",
    tools=[{
        "type": "mcp_server",
        "name": "weather", # Must be lowercase
        "url": "https://gemini-api-demos.uc.r.appspot.com/mcp"
    }]
)

print(interaction.output_text)

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "What is the weather in Tokyo?",
    environment: "remote",
    tools: [{
        type: "mcp_server",
        name: "weather", // Must be lowercase
        url: "https://gemini-api-demos.uc.r.appspot.com/mcp"
    }]
}, { timeout: 300000 });

console.log(interaction.output_text);

REST

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -d '{
      "agent": "antigravity-preview-05-2026",
      "input": "What is the weather in Tokyo?",
      "environment": "remote",
      "tools": [{
          "type": "mcp_server",
          "name": "weather",
          "url": "https://gemini-api-demos.uc.r.appspot.com/mcp"
      }]
  }'

Customizing the agent

You can extend the Antigravity agent by customizing its instructions, tools, and environment. The agent supports a filesystem-native approach to customization: you can mount files like AGENTS.md for instructions and skills under .agents/skills/ directly into the sandbox, or pass configuration inline at interaction time. You can iterate on your configuration inline and then save it as a managed agent when you are ready.

For full details on how to build custom agents, see Building Managed Agents.

Background execution

Agent tasks that involve multi-step reasoning, code execution, or file operations can take minutes to complete. Use background=True to run the interaction asynchronously. The API returns immediately with an interaction ID that you poll until the status is completed or failed.

Python

import time
from google import genai

client = genai.Client()

# 1. Start the interaction in the background
interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="Run a complex analysis on the repository.",
    environment="remote",
    background=True,
)

print(f"Interaction started in background: {interaction.id}")

# 2. Poll for completion
while interaction.status == "in_progress":
    time.sleep(5)
    interaction = client.interactions.get(id=interaction.id)

if interaction.status == "completed":
    print(interaction.output_text)
else:
    print(f"Finished with status: {interaction.status}")

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "Run a complex analysis on the repository.",
    environment: "remote",
    background: true,
});

console.log(`Interaction started in background: ${interaction.id}`);

let result = interaction;
while (result.status === "in_progress") {
    await new Promise(resolve => setTimeout(resolve, 5000));
    result = await client.interactions.get(interaction.id);
}

if (result.status === "completed") {
    console.log(result.output_text);
} else {
    console.log(`Finished with status: ${result.status}`);
}

REST

# 1. Start the interaction in the background
RESPONSE=$(curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Api-Revision: 2026-05-20" \
  -d '{
      "agent": "antigravity-preview-05-2026",
      "input": "Run a complex analysis on the repository.",
      "environment": "remote",
      "background": true
  }')

INTERACTION_ID=$(echo $RESPONSE | jq -r '.id')

# 2. Poll for results (repeat until status is "completed")
curl -s -X GET "https://generativelanguage.googleapis.com/v1beta/interactions/$INTERACTION_ID" \
  -H "x-goog-api-key: $GEMINI_API_KEY"

Background execution requires store=True, which is the default. For real-time progress updates during background execution, see Streaming background interactions.

You can cancel a running background interaction using the cancel method.

Python

client.interactions.cancel(id="INTERACTION_ID")

JavaScript

await client.interactions.cancel({ id: "INTERACTION_ID" });

REST

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions/INTERACTION_ID:cancel" \
  -H "x-goog-api-key: $GEMINI_API_KEY"

Multi-turn with background execution

When a background interaction involves stateful tools (like code execution in a sandbox), use the environment_id from the completed interaction to continue in the same environment. This ensures the agent picks up where it left off with all files and state intact.

Python

import time
from google import genai

client = genai.Client()

# First turn: run a task in the background
interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="Clone https://github.com/google/generative-ai-python and run its tests.",
    environment="remote",
    background=True,
)

while interaction.status == "in_progress":
    time.sleep(5)
    interaction = client.interactions.get(id=interaction.id)

# Second turn: continue in the same environment
followup = client.interactions.create(
    agent="antigravity-preview-05-2026",
    input="Fix any failing tests and re-run them.",
    previous_interaction_id=interaction.id,
    environment=interaction.environment_id,
    background=True,
)

while followup.status == "in_progress":
    time.sleep(5)
    followup = client.interactions.get(id=followup.id)

print(followup.output_text)

JavaScript

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

// First turn: run a task in the background
let interaction = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "Clone https://github.com/google/generative-ai-python and run its tests.",
    environment: "remote",
    background: true,
});

while (interaction.status === "in_progress") {
    await new Promise(resolve => setTimeout(resolve, 5000));
    interaction = await client.interactions.get(interaction.id);
}

// Second turn: continue in the same environment
let followup = await client.interactions.create({
    agent: "antigravity-preview-05-2026",
    input: "Fix any failing tests and re-run them.",
    previous_interaction_id: interaction.id,
    environment: interaction.environment_id,
    background: true,
});

while (followup.status === "in_progress") {
    await new Promise(resolve => setTimeout(resolve, 5000));
    followup = await client.interactions.get(followup.id);
}

console.log(followup.output_text);

REST

# 1. Start first interaction in the background
RESPONSE=$(curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Api-Revision: 2026-05-20" \
  -d '{
      "agent": "antigravity-preview-05-2026",
      "input": "Clone https://github.com/google/generative-ai-python and run its tests.",
      "environment": "remote",
      "background": true
  }')

INTERACTION_ID=$(echo $RESPONSE | jq -r '.id')

# 2. Poll until completed (repeat until status is "completed")
RESULT=$(curl -s -X GET "https://generativelanguage.googleapis.com/v1beta/interactions/$INTERACTION_ID" \
  -H "x-goog-api-key: $GEMINI_API_KEY")

ENVIRONMENT_ID=$(echo $RESULT | jq -r '.environment_id')

# 3. Continue in the same environment
curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Api-Revision: 2026-05-20" \
  -d "{
      \"agent\": \"antigravity-preview-05-2026\",
      \"input\": \"Fix any failing tests and re-run them.\",
      \"previous_interaction_id\": \"$INTERACTION_ID\",
      \"environment\": \"$ENVIRONMENT_ID\",
      \"background\": true
  }"

Environments

Each call creates or reuses a Linux sandbox. The environment parameter takes three forms:

Form Description
"remote" Provision a fresh sandbox with default settings.
"env_abc123" Reuse an existing environment by ID, preserving all files and state.
{...} Full EnvironmentConfig with custom sources and network rules.

See Environments for details on sources (Git, GCS, inline), networking, lifecycle, and resource limits.

Availability and pricing

Antigravity agent is available in preview through the Interactions API in Google AI Studio and the Gemini API.

Pricing follows a pay-as-you-go model based on the underlying Gemini model tokens and the tools the agent uses. Unlike a standard chat request that produces a single output, an Antigravity interaction is an agentic workflow. A single request triggers an autonomous loop of reasoning, tool execution, code running, and file management.

Estimated costs

Costs vary based on task complexity. The agent autonomously determines how many tool calls, code executions, and file operations are needed. The following estimates are based on runs.

Task category Input tokens Output tokens Typical cost
Research & information synthesis 100k–500k 10k–40k $0.30–$1.00
Document & content generation 100k–500k 15k–50k $0.30–$1.30
Process & system design 100k–400k 10k–30k $0.25–$0.80
Data processing & analysis 300k–3M 30k–150k $0.70–$3.25

50–70% of input tokens are typically cached. Complex agentic workflows with many tool calls can accumulate 3–5 million tokens in a single interaction, with costs up to ~$5.

Environment compute (CPU, memory, sandbox execution) is not billed during the preview period.

Limitations

  • Preview status: The Antigravity agent and the Interactions API. Features and schemas may change.
  • Unsupported generation config: The following parameters are not supported and return a 400 error: temperature, top_p, top_k, stop_sequences, max_output_tokens.
  • Structured output: The Antigravity agent does not support structured outputs.
  • Unavailable tools: file_search, computer_use, and google_maps are not yet supported.
  • Remote MCP limitations: Server-Sent Events (SSE) transport is not supported (use Streamable HTTP). Additionally, the server name must be strictly lowercase and alphanumeric (using uppercase letters triggers a generic 400 Bad Request error).
  • Filesystem tool: There is no filesystem tool at the moment. It is part of the environment.
  • Store requirement: Agent execution using background=True requires store=True.
  • Stateful only function calling: Function calling is only supported in stateful mode. You must use previous_interaction_id to continue the turn; reconstructing history manually (stateless mode) is not supported.
  • Unsupported multimodal types. Audio, video, and document inputs are not supported at the moment. Only text and image are allowed.

What's next