跳到主要内容

gemini-media

❖ Communityv0.1.2

Tools · by bytecakelake · added

Google Gemini image (Nano Banana) and video (Gemini Omni) generation backends for image_generate and video_generate, in one plugin built on the google-genai SDK. Image: text-to-image and reference-guided editing (up to 14 images). Video: text-, image- and reference-to-video with native audio. Model picker comes from the key's live models.list.

Open in Hermes Desktop
hermes plugins install gemini-media
Repository ↗Reviewed source @ 44bc2dd ↗Documentation ↗

What it adds

Environment variables it needs 1

GEMINI_API_KEY

README

From the reviewed commit 44bc2dd ↗; it updates when the author re-pins.

hermes-gemini-media-generation

Google Gemini image (Nano Banana) and video (Gemini Omni) backends for Hermes Agent image_generate and video_generate, in a single plugin built on the google-genai SDK.

Status: v0.1.2 (MVP). Image and video generation work against the live Gemini API with an API key. OAuth sign-in is planned for a later release. The Gemini Omni video service stalls intermittently; see Known behaviour.

Features

  • image_generate: Nano Banana models, text-to-image and reference-guided editing (up to 14 reference images), 1K/2K/4K.
  • video_generate: Gemini Omni, text-to-video, image-to-video and reference-to-video with native audio, 3-10 s, 16:9 / 9:16, 360p-4K.
  • Model picker built from your key's live models.list (cached, with an offline fallback).
  • One plugin, one API key, built on the official google-genai SDK.

See docs/MODELS.md for models, measured limits, parameter handling and where generated files are stored.

Install

hermes plugins install bytecakelake/hermes-gemini-media-generation
hermes plugins enable gemini-media
hermes tools        # Image Generation / Video Generation -> Google Gemini

GEMINI_API_KEY (from https://aistudio.google.com/apikey) goes in ~/.hermes/.env. If you set it as a Windows system environment variable, restart Hermes so it is picked up. GOOGLE_API_KEY is accepted as a fallback.

Model selection

The agent never picks a model. Order of precedence:

  1. GEMINI_IMAGE_MODEL / GEMINI_VIDEO_MODEL environment variable
  2. image_gen.gemini.model / video_gen.gemini.model in config.yaml
  3. image_gen.model / video_gen.model (only if it is a Gemini model id)
  4. built-in default (gemini-nano-banana-2.1, gemini-omni-1.1-flash)

The model picker is built from the Gemini models.list endpoint for your key, cached for 24 h in $HERMES_HOME/cache/gemini_media/models.json, with a built-in fallback list when the endpoint is unreachable. The cache stores a hash of the key, never the key.

Development

python -m venv .venv && .venv/Scripts/python -m pip install -e ".[dev]"
.venv/Scripts/python -m pytest                 # offline unit tests
GEMINI_API_KEY=... .venv/Scripts/python -m pytest -m live   # real API calls
hermes plugins validate .

Provider tests subclass Hermes's ABCs, so a Hermes source tree must be importable. The tests look in %LOCALAPPDATA%\hermes\hermes-agent or the directory in HERMES_AGENT_DIR.

Source lives in src/hermes_gemini_media/; the root __init__.py is a three-line shim because Hermes loads plugins from the repository root.

What this plugin does on your machine

Disclosure, as required for Hermes catalog listing:

  • Network: sends your prompt, any reference images and your API key to generativelanguage.googleapis.com (the Gemini API). Reference images given as URLs are downloaded by the plugin first (public addresses only; private/loopback hosts are refused).
  • Files read: only local image files you pass as image_url / reference_image_urls (header-checked: anything that is not a PNG/JPEG/WebP/HEIC/HEIF image is rejected before it is read in full). No credential files, no .env, no auth.json.
  • Files written: generated media under $HERMES_HOME/cache/generated/, and the model list cache $HERMES_HOME/cache/gemini_media/models.json (contains a hash of the key, never the key).
  • Credentials: read through Hermes's get_secret; never logged, and scrubbed from errors.
  • Cost: image and video generation are billed by Google to the API key's project.
  • No telemetry, background processes, shell commands or self-updates.

Known behaviour

  • The Gemini Omni video endpoint sometimes stalls for minutes (observed in October 2026). Calls are bounded by a wall-clock deadline (default 300 s for video, 150 s for images; set video_gen.gemini.timeout_s / image_gen.gemini.timeout_s) and return error_type: timeout. The plugin does not retry automatically because an abandoned request may still be billed.
  • Do not request task: text_to_video explicitly: that value hung every time in testing, so the plugin leaves the task unset for text-only prompts.

Relationship to other plugins

semirkabir/hermes-plugin-gemini-image also provides Gemini image generation. This project is an independent implementation (no code shared) that adds video generation and a live model catalog in one plugin. Both register the provider name gemini; enable only one of them.

License

Apache License 2.0. See LICENSE and NOTICE.

Disclosure

← Back to the catalog · catalog built Oct 25, 2026