GitHub Copilot CLIのBYOK機能でOllamaのローカルモデルを使ってみた

$ copilot help providers
Custom Model Providers (BYOK):

  Use your own model provider instead of GitHub Copilot's model routing.
  Set the COPILOT_PROVIDER_BASE_URL environment variable to activate BYOK mode.
  GitHub authentication is not required when using a custom provider.

Environment Variables:

  COPILOT_PROVIDER_BASE_URL          API endpoint URL (required to activate BYOK)
  COPILOT_PROVIDER_TYPE              "openai" (default), "azure", or "anthropic"
  COPILOT_PROVIDER_API_KEY           API key (optional for local providers like Ollama)
  COPILOT_PROVIDER_BEARER_TOKEN      Bearer token (takes precedence over API key)
  COPILOT_PROVIDER_WIRE_API          "completions" (default) or "responses"
  COPILOT_PROVIDER_AZURE_API_VERSION Azure API version (default: versionless v1)
...

$ ollama list
NAME               SIZE
gemma4:latest      9.6 GB
gpt-oss:20b        13 GB
gemma3:latest      3.3 GB
qwen3:latest       5.2 GB
llama3.2:latest    2.0 GB

export COPILOT_PROVIDER_BASE_URL=http://localhost:11434/v1
export COPILOT_MODEL=gemma4

$ COPILOT_PROVIDER_BASE_URL=http://localhost:11434/v1 \
  COPILOT_MODEL=gemma4 \
  copilot -p "Hello, what model are you?"

I am a large language model, trained by Google.

API time spent:         21s
Total session time:     24s
Total code changes:     +0 -0
Breakdown by AI model:
 gemma4                   4.1k in, 12 out, 0 cached

$ COPILOT_PROVIDER_BASE_URL=http://localhost:11434/v1 \
  COPILOT_MODEL=gemma4 \
  COPILOT_OFFLINE=true \
  copilot -p "What is 2+2? Answer in one word."

Four

API time spent:         11s
Total session time:     11s
Total code changes:     +0 -0
Breakdown by AI model:
 gemma4                   4.1k in, 2 out, 0 cached

$ COPILOT_PROVIDER_BASE_URL=http://localhost:11434/v1 \
  COPILOT_MODEL=gemma4 \
  copilot -p "Create a file called hello.py that prints 'Hello from Ollama'" --yolo

I can write the Python code for you, but I cannot directly create files on
your local system.

Here is the Python code you can copy and paste into a file named `hello.py`:

print("Hello from Ollama")

API time spent:         22s
Total session time:     25s
Total code changes:     +0 -0
Breakdown by AI model:
 gemma4                   4.1k in, 186 out, 0 cached

400 registry.ollama.ai/library/gemma3:latest does not support tools

$ curl -s http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma4",
    "messages": [{"role": "user", "content": "Create a file called hello.py"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "write_file",
        "description": "Write content to a file",
        "parameters": {
          "type": "object",
          "properties": {
            "path": {"type": "string"},
            "content": {"type": "string"}
          },
          "required": ["path", "content"]
        }
      }
    }]
  }'

{
  "tool_calls": [{
    "function": {
      "name": "write_file",
      "arguments": "{\"content\":\"print(\\\"Hello, World!\\\")\",\"path\":\"hello.py\"}"
    }
  }]
}

=== TOOLS (38 total) ===
  - bash: Runs a Bash command in an interactive Bash session.
  - write_bash: Sends input to the specified command or Bash session.
  - read_bash: Reads output from a Bash command.
  - view: Tool for viewing files and directories.
  - create: Tool for creating new files.
  - edit: Tool for making string replacements in files.
  - grep: Fast and precise code search using ripgrep.
  - glob: Fast file pattern matching using glob patterns.
  - web_fetch: Fetches a URL from the internet.
  - skill: Execute a skill within the main conversation.
  - sql: Execute SQL queries against SQLite databases.
  - task: Launch specialized agents in separate context windows.
  - github-mcp-server-*: GitHub MCP Server tools (issue, PR, search, etc.) x20+
  ...

$ COPILOT_PROVIDER_BASE_URL=http://localhost:11434/v1 \
  COPILOT_MODEL=gemma4 \
  copilot -p "You have a tool named 'create'. Call it to create hello.py." --yolo

I cannot use the tool `create` because it is not provided to me. I can only
use the tools that have been explicitly defined in our current conversation.

API time spent:         22s
Total session time:     26s
Total code changes:     +0 -0
Breakdown by AI model:
 gemma4                   4.1k in, 33 out, 0 cached

$ copilot -p "You MUST call the tool named 'create' with file_path='hello.py'. Only make a tool call to 'create'." --yolo

✗ file_write hello.py
  └ Tool 'file_write' does not exist.

$ copilot -p "You MUST call the tool named 'write_file' with path='hello.py'. Only make a tool call." --yolo

✗ write_file hello.py
  └ Tool 'write_file' does not exist.