Create a new Antigravity agent interaction.

Args: prompt: Task instruction or prompt for the agent. agent: Agent identifier (default: antigravity-preview-05-2026). model: Model name for agent_config (e.g. gemini-3.6-flash, gemini-3.5-flash-lite). environment: Sandbox environment mode ('remote' or existing environment_id). tools: List of custom tool dicts (code_execution, google_search, url_context, mcp_server, function). images: Optional list of base64 images, e.g. [{"data": "", "mime_type": "image/png"}]. max_total_tokens: Optional token budget cap for the interaction. background: Run asynchronously; poll with get_interaction_status. Forces store=True. system_instruction: Optional behavior/persona override for this interaction. stream: Stream the response via the SDK (ignored by the REST fallback path).

Continue an existing interaction turn with new prompt, images, or function results.

Args: previous_interaction_id: ID of the interaction turn to continue. environment_id: Sandbox environment ID returned from previous turn. prompt: Optional follow-up prompt text. images: Optional follow-up base64 images. function_results: Optional list of function_result dicts: [{"name": "...", "call_id": "...", "result": {...}}]. agent: Agent identifier. system_instruction: Optional behavior/persona override for this turn.

Submit a local function execution result back to a pending interaction (requires_action status).

Get status and current steps of an interaction.

Cancel an in-progress (e.g. background) interaction.

Delete a stored interaction.

List cached interaction sessions and their current status, newest first.

Args: page_size: Max sessions to return (default 50). page_token: Offset into the cache to resume from; pass back the returned "next_page_token".

Helper tool to generate remote MCP server configuration dict for agent tools parameter.

Helper tool to generate a remote sandbox EnvironmentConfig dict.

Args: sources: List of {"type": "repository"|"gcs"|"inline", "source": "url-or-content", "target": "/workspace/path", "content": "..."}. "repository" (git URL) and "gcs" use "source"; "inline" uses "content" instead. "target" is required for all and must be a subdirectory (cannot mount to "/"). network_allowlist: List of {"domain": "api.example.com"|".example.com"|"", "transform": {"Authorization": "Bearer ..."}}. Ignored if network_disabled is True. network_disabled: Block all outbound network traffic (sets network="disabled"). environment_id: Reuse an existing environment's state instead of provisioning fresh.

Build a hooks.json config for sandbox-native tool interception (code_execution, read/write/list/delete_file).

Hooks only fire for sandbox-native tools, never for function calling or mcp_server tools. Mount the result at target ".agents/hooks.json" via build_environment_config's inline sources so the sandbox picks it up automatically.

Args: pre_tool_execution: List of {"matcher": "<RE2 regex, e.g. 'code_execution' or '.*_file'>", "hooks": [{"type": "command"|"http", "command"|"url": "...", "timeout": 10}]}. A "command" hook receives the tool call as JSON on stdin and must print {"decision": "allow"} or {"decision": "deny", "reason": "..."} to stdout. post_tool_execution: Same matcher/hooks shape; runs after execution, cannot block, output ignored. group_name: Top-level key grouping this hook set. enabled: Whether this group is active.

Have the agent clone a GitHub repo, branch off base_branch, do the task, and open a PR.

Token is injected via the sandbox network allowlist header transform (never placed in the prompt text), so it stays out of interaction transcripts/logs.

Args: repo_url: e.g. "https://github.com/owner/repo" (public or private). task_prompt: What the agent should implement/fix. base_branch: Branch to start from (default "main"). work_branch: Branch name to create; auto-generated if omitted. pr_title: PR title; defaults to a summary derived from task_prompt. github_token: GitHub PAT with repo scope. Falls back to GITHUB_TOKEN env var. model: Gemini model for the interaction. agent: Agent identifier.

Create a cron-scheduled trigger that runs an Antigravity interaction on a recurring schedule.

Args: schedule: Cron expression, e.g. "0 9 * * *". prompt: Task instruction the agent runs on each execution. time_zone: IANA time zone, e.g. "America/Argentina/Buenos_Aires". display_name: Human-readable trigger name. agent: Agent identifier. environment: Sandbox environment mode ('remote' or existing environment_id). max_consecutive_failures: Auto-pause trigger after N consecutive failures. execution_timeout_seconds: Per-execution timeout.

List all cron triggers.

Fetch a single trigger by ID.

Pause or resume a trigger. status: 'active' or 'paused'.

Permanently delete a trigger.

Execute a trigger immediately, outside its schedule.

View execution history for a trigger.

Register a reusable named agent (invoke later via create_interaction(agent=agent_id)).

Args: agent_id: Unique identifier. Cannot start with reserved prefixes (antigravity-, gemini-, google-, etc). model: Model for agent_config (default gemini-3.6-flash). system_instruction: Persistent persona/behavior definition. tools: Default tools; falls back to code_execution, google_search, url_context if omitted. base_environment: Environment string/id or config dict (see build_environment_config); forked per invocation. description: Human-readable description.

List all registered managed agents in the project.

Fetch a managed agent's stored configuration.

Delete a managed agent's registration (existing environments/interactions persist).

Provision a standalone sandbox environment for reuse across multiple interactions.

Unlike passing environment=build_environment_config(...) inline to create_interaction (which forks per invocation), this persists an environment_id you can pass to create_interaction/continue_interaction directly.

Args: sources: See build_environment_config. network_allowlist: See build_environment_config. network_disabled: See build_environment_config.

List sandbox environments in the project.

Fetch a sandbox environment's details (type, sources, network config).

Terminate and remove a sandbox environment's resources.

Download the full sandbox filesystem as a .tar archive to a local path.

Args: environment_id: Environment to snapshot. dest_path: Local filesystem path to write the .tar archive to.