Antigravity CLI Migration API Guide (2026): What Gemini CLI Users Should Change Now

Google’s May 19, 2026 developer announcement has a hard date developers shouldn’t ignore: on June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions stop serving requests for Google AI Pro and Ultra users, plus free Gemini Code Assist for individuals. Google’s recommended path is Antigravity CLI and Antigravity 2.0.

The headline sounds like a product migration. In practice, it’s a workflow migration. If your team used Gemini CLI as a cheap terminal agent, a PR helper, a refactor bot, or a background research runner, the important question isn’t “what command do I install next?” It’s “how do I keep coding agents reliable, measurable, and portable when the default surface changes?”

This guide gives you a practical migration plan. It focuses on the parts that tend to break: API keys, model routing, token budgets, background jobs, and fallback paths.

What Google Actually Announced

According to Google’s developer blog, Antigravity CLI is now available to everyone. Google also says Antigravity CLI keeps several important Gemini CLI concepts: Agent Skills, Hooks, Subagents, and Extensions, now framed as Antigravity plugins. The new pitch is faster execution, asynchronous multi-agent workflows, and a unified agent harness shared with Antigravity 2.0.

The consumer timeline matters:

That split is easy to miss. If you’re enterprise, don’t panic. If you’re on individual access, don’t wait until a CI job or a weekend refactor suddenly stops working.

Migration Step 1: Inventory How Gemini CLI Is Used

Start with usage, not tooling. Most teams underestimate how many little scripts depend on a coding agent.

rg -n "gemini|gemini-cli|code assist|GEMINI_API_KEY|GOOGLE_API_KEY" \
  .github scripts package.json pnpm-lock.yaml yarn.lock README.md

Group each hit into one of four buckets:

• Interactive terminal work: local prompts, one-off debugging, repo explanation.
• IDE workflows: inline code changes, refactors, commit help.
• CI or GitHub review: PR comments, test summaries, release notes.
• Background agents: long-running research, multi-file edits, dependency upgrades.

The third and fourth buckets need the most care. They spend tokens while you’re not watching.

Migration Step 2: Put a Routing Layer Between Tools and Models

Hard-coding one provider into every agent script is the fastest way to repeat this migration pain later. A thin router lets Antigravity CLI, Claude Code, Codex CLI, or your own scripts call a stable internal endpoint while you decide which upstream model handles each task.

Here’s a simple OpenAI-compatible request shape you can use as the common contract:

curl https://api.kissapi.ai/v1/chat/completions \
  -H "Authorization: Bearer $KISSAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.5-flash",
    "messages": [
      {"role": "system", "content": "You are a careful coding agent. Return a concise patch plan first."},
      {"role": "user", "content": "Review this pull request for risky database changes."}
    ],
    "temperature": 0.2
  }'

KissAPI is useful here because it gives you one OpenAI-compatible surface for multiple model families. That doesn’t replace Antigravity CLI; it makes your model access less brittle around it.

Migration Step 3: Separate Interactive Runs from Background Runs

Antigravity CLI’s asynchronous workflow direction is good. It also makes cost control more important. Background agents can quietly do expensive things: read the same files repeatedly, spawn subagents, retry failed commands, and produce huge summaries.

Give each run type a budget before you migrate:

Before you move a job, estimate it. KissAPI’s token counter is handy for checking prompt size, and the API cost calculator is better than guessing from vibes.

Migration Step 4: Add a Small Node.js Compatibility Wrapper

If your existing scripts assume a Gemini-style provider but you want portable routing, hide provider choice behind a tiny client. This example uses an OpenAI-compatible endpoint so the rest of your script doesn’t care what model sits behind it.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.KISSAPI_API_KEY,
  baseURL: process.env.AI_BASE_URL || "https://api.kissapi.ai/v1"
});

export async function askCodingAgent({ model, task, files }) {
  const response = await client.chat.completions.create({
    model,
    temperature: 0.2,
    messages: [
      {
        role: "system",
        content: "You are a senior coding agent. Be specific, avoid broad rewrites, and flag risky assumptions."
      },
      {
        role: "user",
        content: `Task:\n${task}\n\nFiles:\n${files.join("\n")}`
      }
    ]
  });

  return response.choices[0].message.content;
}

Then your CI job can switch models with an environment variable instead of a code change:

AI_BASE_URL="https://api.kissapi.ai/v1" \
KISSAPI_API_KEY="..." \
node scripts/review-pr.js --model gemini-3.5-flash

Migration Step 5: Keep Fallback Simple

Don’t build a giant orchestration system if all you need is “try model A, then model B.” A boring fallback is easier to debug at 2 a.m.

MODELS=("gemini-3.5-flash" "claude-sonnet-4-6" "gpt-5.5-mini")

for MODEL in "${MODELS[@]}"; do
  echo "Trying $MODEL"
  if node scripts/run-agent.js --model "$MODEL"; then
    exit 0
  fi
  sleep 3
done

echo "All agent routes failed" >&2
exit 1

The right fallback depends on the task. For long code edits, compare model behavior on a fixed evaluation set before switching traffic. For quick explanations and summaries, failover can be more aggressive.

What Not to Migrate Blindly

Be careful with these three things:

• Plugin permissions: Any tool that can read files, run shell commands, or open network connections needs review after migration.
• Hidden context inflation: Multi-agent systems may pass large state between workers. Measure actual tokens, not just the user prompt.
• CI comments: Automated PR review is public inside your company. Keep secrets, credentials, and private business logic out of prompts.

My opinion: this transition is a warning shot for every team leaning on plan-bundled AI access. Coding agents are becoming real infrastructure. Treat them like infrastructure. Put them behind budgets, logs, routing, and permission boundaries.

FAQ

What changed for Gemini CLI users on June 18, 2026?

Google said Gemini CLI and Gemini Code Assist IDE extensions would stop serving requests for Google AI Pro, Ultra, and free individual Code Assist users on June 18, 2026. Antigravity CLI is the recommended consumer migration path.

Do enterprise Gemini Code Assist users need to migrate immediately?

No. Google stated that Standard, Enterprise, Google Cloud GitHub integrations, and paid API-key users keep access unchanged. Still, it’s smart to test Antigravity CLI because future agent features will likely land there first.

How do I control API costs while moving coding agents?

Track token use by workflow, set per-run limits, route easy tasks to cheaper models, and keep fallback models for outages. Use the token counter before large runs and the cost calculator before enabling background agents.