Google Antigravity release notes: CLI 1.1.1 selects agents and SDK 0.1.6 opens local models

Google Antigravity picked up two useful updates since the previous release-notes review. CLI 1.1.1 can select a custom agent at launch, while Python SDK 0.1.6 adds configurations for local LiteRT models and external OpenAI-compatible model servers. For teams that want to reuse specialised agents, that matters more than another long list of fixes.

Google Antigravity is Google's agentic development environment and SDK for AI agents that can read projects, use tools, and complete multi-step jobs. A custom agent is a reusable role with its own instructions and tools. A local model runtime runs the model through a local runtime or model server instead of a standard cloud model.

Google Antigravity release notes: CLI 1.1.1 selects an agent at launch

Antigravity CLI 1.1.1 was released on July 10, 2026. It adds an --agent flag for selecting a custom agent at launch and an agent/agents subcommand for listing available agents. The release notes name the feature but do not show the complete command syntax, so check the CLI help in your installation before putting it into an automated launch command.

This makes role selection part of the run itself. A team can keep separate agents for release checks, data migrations, and documentation without opening every session with the same long role description.

The release also fixes two issues that matter in automation. --print/-p now writes server-side failures to stderr and exits with a non-zero code, while agy -p should no longer hang in scripts when the prompt was already supplied through a flag. Scheduled runs should be easier to monitor as a result.

Human step: update to Antigravity CLI 1.1.1, use the CLI help to confirm the exact syntax, then select an existing custom agent with --agent. If you have pre-approved write_file in settings.json under permission.allow, note that default mode now respects that rule and will not ask for review on every such write.

Source: Antigravity CLI 1.1.1 on GitHub

SDK 0.1.6 opens more local model paths

The google-antigravity Python package 0.1.6 was published on PyPI on July 9, 2026. The public code at that tag exports two new configurations: LiteRTAgentConfig for LiteRT and LocalOpenAIAgentConfig for external OpenAI-compatible APIs. Its code documentation names Ollama and LM Studio as examples of local model servers.

This is a developer feature in a package still labelled Alpha. The practical use is clear enough: teams can start testing the same agent framework through more local model paths. Workspaces, tools, MCP servers, subagents, and policies can stay around the agent while the model connection changes. MCP, or Model Context Protocol, is a standard way to connect AI agents to external tools and data sources.

The SDK code enables file writes, command execution, subagents, and MCP as default capabilities in LocalOpenAIAgentConfig. Tie the technical test to scoped workspaces, approval points, and run logs. Keep credentials in environment variables or a secret manager rather than in the prompt.

Sources: google-antigravity 0.1.6 on PyPI, new SDK exports in tag v0.1.6, LocalOpenAIAgentConfig in v0.1.6

What Nordic teams should test now

Start with one focused role that already has a clear review job, such as a release checker. Select that role when starting the CLI and run the same read-only check against a known repository. See whether agent selection really shortens setup and whether your script catches failures through the exit code and stderr.

If you build Python agent flows, test SDK 0.1.6 separately against a local model server. Compare answer quality, latency, tool calls, and failure handling with your current model path before enabling writes. This is a natural Tool Forge job: connect the agent, model, and tools so every run can be reviewed.

Short example: use the new Antigravity feature

Run this prompt after you, as the human operator, have selected your release agent at launch. It contains no installation or interface steps.

Act as our release checker for this repository.
Read the latest changes and relevant test results.
Return risk areas, files to review, and a read-only validation plan.
Change nothing until I approve the plan.

Good output should include:

• concrete risk areas tied to the latest changes
• filenames and a short reason for each review choice
• a validation plan that does not write to the repository
• a clear stopping point before any changes

What to watch next

The app's public changelog still stops at Antigravity 2.2.1 from June 25. The next things to watch are fuller CLI documentation for custom agent selection and an official guide for the SDK's new local model configurations. Until then, treat 0.1.6 as a solid developer signal, not a finished migration recipe.

Source: Google Antigravity changelog