llm-gemini-code-assist 0.28.2

LLM plugin to access Google's Gemini via Code Assist API with OAuth

llm-gemini-code-assist

API access to Google's Gemini models via the Code Assist API with OAuth authentication.

This is a fork of llm-gemini modified to use Google's Code Assist API, which requires OAuth authentication instead of API keys.

Installation

Install this plugin in the same environment as LLM:

llm install llm-gemini-code-assist

Authentication

Unlike the standard llm-gemini plugin, this version uses OAuth authentication with the Code Assist API. Authentication generally is automatic if you have gemini-cli installed; else, you can authenticate manually:

llm gemini-ca auth

This will:

1. Open your browser to Google's OAuth consent page
2. After you approve, save credentials to ~/.gemini/oauth_creds.json
3. The credentials include a refresh token for automatic renewal

The OAuth credentials are stored with file permissions 0600 for security.

Usage

Once authenticated, use the models with the gemini-ca/ prefix:

llm -m gemini-ca/gemini-2.5-flash "Tell me a joke about a pelican"

You can set it as your default model:

llm models default gemini-ca/gemini-2.5-flash
llm "Tell me a joke about a pelican"

Available Models

Models available via Code Assist API with the gemini-ca/ prefix:

• gemini-ca/gemini-3-pro-preview - Gemini 3 Pro (preview)
• gemini-ca/gemini-3-flash-preview - Gemini 3 Flash (preview)
• gemini-ca/gemini-2.5-pro - Gemini 2.5 Pro
• gemini-ca/gemini-2.5-flash - Gemini 2.5 Flash
• gemini-ca/gemini-2.5-flash-lite - Gemini 2.5 Flash Lite

See gemini-cli model configs for the canonical model list.

Model Latency

Typical response times for a simple prompt (tested January 2026):

Model	Avg (ms)	Notes
gemini-2.5-flash-lite	~1200	Fastest
gemini-2.5-flash	~1500
gemini-3-flash-preview	~1900	Preview
gemini-2.5-pro	~4000	High variance
gemini-3-pro-preview	~4500	Preview

Note: Code Assist API has rate limits (~3 requests/minute). Use scripts/profile_request.py to benchmark.

Features

All features from llm-gemini are supported:

Multi-modal Input

llm -m gemini-ca/gemini-2.5-flash 'describe image' -a image.jpg

JSON Output

llm -m gemini-ca/gemini-2.5-flash -o json_object 1 \
  '3 largest cities in California'

Code Execution

llm -m gemini-ca/gemini-2.0-flash -o code_execution 1 \
  'calculate factorial of 13'

Google Search

llm -m gemini-ca/gemini-2.5-flash -o google_search 1 \
  'What happened today?'

Chat

llm chat -m gemini-ca/gemini-2.5-flash

Troubleshooting

If you get authentication errors:

1. Check if your credentials are expired:

   cat ~/.gemini/oauth_creds.json | python -m json.tool
   

2. Re-authenticate:

   llm gemini-ca auth
   

Development

To set up the development environment:

cd llm-gemini-code-assist
uv sync
uv run pre-commit install

This installs dependencies and sets up pre-commit hooks (including secret scanning).

Run tests:

uv run pytest

The pre-commit hooks will automatically run linting, formatting, type checking, and secret scanning before each commit. You can also run them manually:

uv run pre-commit run --all-files

Releasing

Releases are tag-driven and automated via GitHub Actions.

# 1. Ensure tests pass
uv run pytest

# 2. Create and push a version tag (triggers release workflow)
git tag v0.27.0
git push origin v0.27.0

This will automatically build the package, create a GitHub Release with changelog, and publish to PyPI.

Differences from llm-gemini

• Uses OAuth authentication instead of API keys
• Requires Code Assist API access
• Models use gemini-ca/ prefix
• Tokens auto-refresh using stored refresh tokens

License

Apache 2.0