Gemini CLI Setup

Gemini CLI is Google's AI coding assistant and is useful for large-context code tasks.

Complete the Environment Setup chapter first and make sure Node.js and npm are installed.

Linux / macOS

Step 1: Install Gemini CLI

npm i -g @google/gemini-cli

Step 2: Configure Environment Variables

Add to ~/.bashrc or ~/.zshrc:

# API base URL: domain only, do not append /v1beta
export GOOGLE_GEMINI_BASE_URL="https://api.icodeeasy.cc"

# Use your I Code Easy API Key
export GEMINI_API_KEY="your API Key"

# Recommended fast model. You can also use stronger models such as gemini-3.1-pro-preview
export GEMINI_MODEL="gemini-3.8-flash"

# Explicitly use Gemini API v1beta
export GOOGLE_GENAI_API_VERSION="v1beta"

Save the file and run source ~/.bashrc or source ~/.zshrc.

Note: Do not set GOOGLE_GEMINI_BASE_URL to https://api.icodeeasy.cc/v1beta, or Gemini CLI will build paths like /v1beta/v1beta/models/.... If api.icodeeasy.cc is slow from your network, change GOOGLE_GEMINI_BASE_URL to https://jp.icodeeasy.cc or https://sg.icodeeasy.cc, still without /v1beta.

Step 3: Start Gemini

cd your-project-folder
gemini

Windows

Step 1: Install Gemini CLI

npm i -g @google/gemini-cli

Step 2: Configure Environment Variables in PowerShell

# API base URL: domain only, do not append /v1beta
[Environment]::SetEnvironmentVariable("GOOGLE_GEMINI_BASE_URL", "https://api.icodeeasy.cc", "User")

# Use your I Code Easy API Key
[Environment]::SetEnvironmentVariable("GEMINI_API_KEY", "your API Key", "User")

# Recommended fast model. You can also use stronger models such as gemini-3.1-pro-preview
[Environment]::SetEnvironmentVariable("GEMINI_MODEL", "gemini-3.8-flash", "User")

# Explicitly use Gemini API v1beta
[Environment]::SetEnvironmentVariable("GOOGLE_GENAI_API_VERSION", "v1beta", "User")

Open a new PowerShell window after setting environment variables.

If api.icodeeasy.cc is slow from your network, change GOOGLE_GEMINI_BASE_URL to https://jp.icodeeasy.cc or https://sg.icodeeasy.cc, still without /v1beta.

Step 3: Start Gemini

Open a new PowerShell window, enter your project directory, and start:

cd your-project-folder
gemini

Model Selection

Common models:

Gemini CLI's built-in flash option may resolve to the older gemini-3-flash-preview. If a client or plugin sends gemini-3-flash, this service maps it to gemini-3-flash-preview automatically. The text model gemini-2.5-flash is also available: as of 2026-08-03 it is served under its own name (no longer transparently upgraded to gemini-3.6-flash) and billed at the 2.5 Flash price. As of 2026-10-13, gemini-3.7-flash, gemini-3.6-flash, and gemini-3.5-flash are deprecated by Google: requests sent under these three names are still accepted here, but they are transparently served by gemini-3.8-flash, the model name in responses is gemini-3.8-flash, and billing follows the 3.8 Flash price.

Raw API Test Example

To test the native Gemini API directly without the CLI:

curl "https://api.icodeeasy.cc/v1beta/models/gemini-3.8-flash:streamGenerateContent?alt=sse" \
  -H "x-goog-api-key: your API Key" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [
      {
        "role": "user",
        "parts": [{"text": "Hello, briefly introduce yourself"}]
      }
    ]
  }'

Non-streaming endpoint path:

/v1beta/models/{model}:generateContent

Streaming endpoint path:

/v1beta/models/{model}:streamGenerateContent?alt=sse

Video Analysis Demo

Gemini models support video understanding: send the video as inlineData (base64) together with the request, and the model watches it and returns an analysis in the format you ask for. Common video-capable models: gemini-3.8-flash.

The following demo is complete and runnable: read a local MP4 → build the request → call the non-streaming endpoint → extract and validate the JSON result.

# 0) Set your I Code Easy API Key first
export ICODEEASY_API_KEY="your API Key"

# 1) base64-encode the local video (macOS: base64 -i demo.mp4 | tr -d '\n' > demo.b64)
base64 -w0 demo.mp4 > demo.b64

# 2) Build the request with jq to avoid hand-writing a huge JSON body
jq -n --rawfile data demo.b64 '{
  contents: [
    {
      role: "user",
      parts: [
        {text: "Watch this video and return a summary, per-scene descriptions, and keywords as JSON that strictly follows the schema."},
        {inlineData: {mimeType: "video/mp4", data: $data}}
      ]
    }
  ],
  generationConfig: {
    temperature: 0.1,
    responseMimeType: "application/json",
    responseJsonSchema: {
      type: "object",
      properties: {
        summary: {type: "string"},
        scenes: {type: "array", items: {type: "string"}},
        keywords: {type: "array", items: {type: "string"}}
      },
      required: ["summary", "scenes", "keywords"],
      additionalProperties: false
    }
  }
}' > video-request.json

# 3) Non-streaming call
curl --silent --show-error \
  "https://api.icodeeasy.cc/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: ${ICODEEASY_API_KEY}" \
  -H "Content-Type: application/json" \
  --data-binary @video-request.json \
  --output video-response.json

# 4) Check the finish state and extract the business JSON
jq -e '.candidates[0].finishReason == "STOP"' video-response.json > /dev/null
jq -r '[.candidates[0].content.parts[]? | select(.thought != true) | (.text // "")] | join("")' \
  video-response.json > video-result.json
jq -e . video-result.json

You can confirm the video was recognized (promptTokensDetails shows a VIDEO modality token count):

jq '.usageMetadata' video-response.json

Notes:

• Video is billed as input tokens at the same rate as text; the VIDEO modality token count in usageMetadata is the billing basis.
• base64 is about 1/3 larger than the original file; keep each clip to about 10–15 MB and split longer videos into segments.
• Passing an external URL via fileData.fileUri is not supported: the model service does not fetch external links. Always download the video locally and send it as inlineData as shown in the demo above.
• Latency grows with clip length, so raise the client timeout (5 minutes or more recommended). Occasional 503 / "service temporarily unavailable" responses can happen at peak times; retry after a few tens of seconds.
• This endpoint is for video understanding/analysis; video generation is not included.

FAQ