GEMINI_API_KEY is not used

[marek-veber]

(Please ensure you are running the latest version of Cloud Cloud IntelliJ with Help > Check for Updates.)

(screenshots are often helpful)

Expected Behavior

Work with api-key

Actual Behavior

I'm not able to specify api-key

Additional Information

• IDE type (e.g. IntelliJ, Pycharm): GoLand
• IDE version: 2025.1.3
• Cloud Code version (Settings > Plugins > Cloud Code): 1.21.0-251
• Cloud SDK (Settings > Cloud Code > Cloud SDK)
  • Are you allowing the plugin to manage the Cloud SDK:
  • Version of the Cloud SDK:
• Skaffold version (If you are using Kubernetes features; Settings > Cloud Code > Kubernetes):
• Operating System: Linux

Feel free to deviate from this template as needed, especially if you are submitting a feature request.

[marek-veber]

In the log I see:

POST https://cloudcode-pa.googleapis.com/v1internal:generateChat
{
  "code": 403,
  "details": [
    {
      "@type": "type.googleapis.com/google.rpc.ErrorInfo",
      "reason": "RESOURCE_USAGE_RESTRICTION_VIOLATED",
      "domain": "googleapis.com",
      "metadata": {
        "service": "cloudaicompanion.googleapis.com",
        "consumer": "projects/gemini-gcp-capi, projects/gemini-gcp-capi"
      }
    },
    {
      "@type": "type.googleapis.com/google.rpc.LocalizedMessage",
      "locale": "en-US",
      "message": "Request is disallowed by organization's constraints/gcp.restrictServiceUsage constraint for 'projects/gemini-gcp-capi, projects/gemini-gcp-capi' attempting to use service 'cloudaicompanion.googleapis.com'."
    }
  ],
  "errors": [
    {
      "domain": "global",
      "message": "Request is disallowed by organization's constraints/gcp.restrictServiceUsage constraint for 'projects/gemini-gcp-capi, projects/gemini-gcp-capi' attempting to use service 'cloudaicompanion.googleapis.com'.",
      "reason": "forbidden"
    }
  ],
  "message": "Request is disallowed by organization's constraints/gcp.restrictServiceUsage constraint for 'projects/gemini-gcp-capi, projects/gemini-gcp-capi' attempting to use service 'cloudaicompanion.googleapis.com'.",
  "status": "PERMISSION_DENIED"
})

[marek-veber]

My access to Gemini is working using GEMINI_API_KEY:

curl   -H 'Content-Type: application/json'   -d '{"contents":[{"parts":[{"text":"Explain how AI works"}]}]}'   -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-pro-preview-05-06:generateContent?key=$GEMINI_API_KEY"
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Of course! Explaining how AI works can seem complicated, but the core idea is surprisingly intuitive. Let's break it down using analogies and simple steps.\n\n### The Core Idea: Learning from Experience\n\nAt its heart, **AI is about creating systems that can learn from data to find patterns and make decisions, rather than being explicitly programmed for every single task.**\n\nThink about how a child learns to recognize a cat. You don't give a child a long list of rules like \"a cat has pointy ears, whiskers, fur, and four legs.\" Instead, you just show them cats.\n\n*   You point and say, \"That's a cat.\"\n*   You show them another one, maybe a different color: \"That's also a cat.\"\n*   You show them a dog and say, \"That is *not* a cat.\"\n\nOver time, the child's brain automatically learns the underlying patterns and features of what makes a \"cat.\" AI works in a very similar way, but it uses data and mathematics instead of a biological brain.\n\n---\n\n### The Three Key Ingredients of Modern AI\n\nTo make this learning process happen, AI needs three things:\n\n1.  **Data (The \"Experience\"):** Lots and lots of it. For an AI to learn what a cat is, it needs to see thousands or even millions of pictures of cats, all labeled \"cat.\" It also needs examples of things that are *not* cats.\n2.  **Algorithm / Model (The \"Brain\"):** This is the framework that processes the data. The most common type of model today is a **Neural Network**, which is loosely inspired by the human brain.\n3.  **Computation (The \"Power\"):** This is the powerful computer hardware (like GPUs) needed to process all the data and run the complex calculations of the model. This is like the energy the brain needs to function.\n\n---\n\n### The Learning Process: Training a Model\n\nLet's stick with our \"Cat vs. Dog\" example. Here’s how an AI model learns, a process called **\"training\"**:\n\n**Step 1: The Initial Guess**\nYou give the untrained AI model (the \"brain\") a picture of a cat. Since it's untrained, its internal \"wiring\" is random. It makes a wild guess.\n\u003e **AI:** \"I'm 55% sure that's a... dog.\"\n\n**Step 2: The Feedback**\nYou compare the AI's guess to the correct label (\"cat\"). The AI was wrong. You calculate how *wrong* it was (this is called the \"error\" or \"loss\").\n\u003e **You:** \"Nope, that was a cat. You were wrong.\"\n\n**Step 3: The Adjustment**\nThis is the magic part. The algorithm uses a mathematical process (called **backpropagation** and **gradient descent**) to slightly adjust its internal connections. Think of it as tuning millions of tiny little knobs.\n*   The connections that led to the wrong answer (\"dog\") are weakened.\n*   The connections that might have led to the right answer (\"cat\") are strengthened.\n\u003e **AI:** \"Okay, I'll adjust my internal settings so I'm a little less likely to say 'dog' for a picture like that in the future.\"\n\n**Step 4: Repeat. A Lot.**\nYou repeat this process millions of times with millions of different pictures of cats and dogs. With each cycle of **Guess -\u003e Check -\u003e Adjust**, the AI gets a tiny bit better.\n\nOver time, after seeing countless examples, the AI model becomes incredibly accurate at identifying the subtle patterns—the ear shapes, nose textures, and eye spacing—that differentiate cats from dogs.\n\n---\n\n### A Look Inside the \"Brain\": The Neural Network\n\nSo what are those \"internal connections\" it's adjusting? In a Neural Network, they are layers of artificial \"neurons.\"\n\n*   **Input Layer:** Takes in the raw data (e.g., the individual pixels of an image).\n*   **Hidden Layers:** This is where the real thinking happens. Each layer recognizes increasingly complex patterns.\n    *   *The first layer* might learn to spot simple edges and colors.\n    *   *The next layer* might combine those edges to recognize shapes like ears and eyes.\n    *   *A deeper layer* might combine those shapes to recognize whole faces or body parts.\n*   **Output Layer:** Gives the final decision (e.g., a probability score: \"98% Cat, 2% Dog\").\n\nThe \"knobs\" the AI tunes are called **weights**, which determine how much influence one neuron has on another. Training is all about finding the perfect set of weights to solve the problem.\n\n---\n\n### What About Generative AI (like ChatGPT and Midjourney)?\n\nThis is a more advanced type of AI, but it's built on the same principles.\n\nInstead of just learning to *label* data (classification), **Generative AI learns the underlying structure and patterns of the data so well that it can *create* new, original data.**\n\n*   **ChatGPT** was trained on a massive amount of text and books from the internet. It didn't just learn to identify nouns and verbs. It learned the relationships between words, sentence structures, context, facts, and even writing styles. When you give it a prompt, it predicts the most probable \"next word\" over and over again to generate coherent sentences and paragraphs.\n\n*   **Midjourney** was trained on a huge dataset of images and their text descriptions. It learned the connection between words like \"astronaut in a photorealistic style\" and the pixel patterns that make up such an image. When you give it a prompt, it starts with digital \"noise\" and gradually refines it to match the patterns it learned from its training data.\n\n### Summary: AI in a Nutshell\n\n| Concept | Simple Analogy | Technical Term |\n| :--- | :--- | :--- |\n| **The Goal** | Learning from experience, not rules. | Pattern Recognition in Data |\n| **The \"Brain\"** | A network of interconnected neurons. | Algorithm / Model (e.g., Neural Network) |\n| **The Learning Process** | Guessing, checking, and getting better. | Training (using backpropagation) |\n| **The \"Knobs\"** | The strength of connections in the brain. | Weights and Biases |\n| **The Fuel** | The books, pictures, and life experience. | Data and Computational Power |\n\nSo, when you ask an AI a question or have it generate an image, you're not talking to a magical thinking machine. You're interacting with a highly complex mathematical model that has been trained on an enormous amount of data to be an expert pattern-matcher and predictor."
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 4,
    "candidatesTokenCount": 1441,
    "totalTokenCount": 2876,
    "promptTokensDetails": [
      {
        "modality": "TEXT",
        "tokenCount": 4
      }
    ],
    "thoughtsTokenCount": 1431
  },
  "modelVersion": "models/gemini-2.5-pro",
  "responseId": "ab9waNGeA5qavdIP9IfmGQ"
}

[ivanporty]

We are not yet supporting Gemini API key for the Gemini Code Assist IDE extensions. We'll update if this changes soon.