The gemini-omni-flash-api Skill: Real Omni Flash Video for Your Agent

What the Skill Is

The gemini-omni-flash-api skill is an official agent skill from Google that teaches a coding agent to drive Gemini Omni Flash, Google's video model, by running real Python scripts instead of hand-writing API calls. It ships in the google-gemini/gemini-skills repository, and you install it with one command.

npx skills add google-gemini/gemini-skills --skill gemini-omni-flash-api

Omni Flash (model ID gemini-omni-flash-preview) launched to developers on June 30, 2026 as a natively multimodal video model in public preview. It ingests any mix of text, image, and video and returns an MP4 with audio through a conversational interface. The problem for an autonomous agent is that Omni has no simple one-shot endpoint, so an agent left to guess the request shape tends to hallucinate parameters. This skill removes the guessing: it hands the agent tested scripts and a SKILL.md that documents exactly how to call them.

What It Wraps

Under the hood the skill wraps two things: the google-genai SDK calls that reach the Interactions API, and ffmpeg preprocessing. Omni Flash runs only through the Interactions API (client.interactions.create), so every generation, edit, and file upload is an interaction. Rather than trusting the model to remember that, the skill encodes it.

import base64
from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-omni-flash-preview",
    input="A marble rolling fast on a chain reaction style track, continuous smooth shot.",
)
with open("marble.mp4", "wb") as f:
    f.write(base64.b64decode(interaction.output_video.data))

That is the call the model would otherwise have to reconstruct from memory on every turn: the exact model ID, the input schema, and how to decode the base64 video out of the response. The script does it the same way every time. For the full API surface behind these scripts, see our Gemini Omni Flash developer guide.

Requirements

Three things must be in place before the skill works: google-genai 2.10.0 or newer, Python 3.10 or newer, and both ffmpeg and ffprobe on your PATH. A Gemini API key is the fourth.

python3 --version                 # need 3.10 or newer
pip install "google-genai>=2.10.0"
ffmpeg -version                   # must be on PATH
ffprobe -version                  # must be on PATH
export GEMINI_API_KEY="your-key"  # from aistudio.google.com/apikey

The version floor matters. The Interactions API itself works from google-genai 1.55.0, but this skill recommends 2.10.0 or newer. If your agent reports that client.interactions does not exist, the SDK is too old. Upgrade before you debug anything else.

The Bundled Scripts

The skill ships four scripts under scripts/video. Each does one job so the agent can compose them.

The example commands the SKILL.md documents look like this:

# Text to video
./scripts/video/generate_video.py "A close-up of a cat drinking tea" \
  --output media/cat_tea.mp4

# Edit an uploaded clip, drop its audio
./scripts/video/generate_video.py "Transform the style to Japanese anime" \
  --video input.mp4 --strip-audio --output media/anime.mp4

# Continue a stored session with no re-upload
./scripts/video/generate_video.py "Change the setting to a snowy winter wonderland." \
  --previous-interaction-id "abc123..." --output media/winter.mp4

Note the pattern in the third command: --previous-interaction-id continues a stored session, so the model edits the previous clip without re-uploading anything. That is the conversational edit chain, exposed as a flag.

Inside Claude Code and Antigravity

Because it is a standard skill, a SKILL.md plus scripts, any agent that can read that file and run shell commands can use it. That includes Anthropic's Claude Code and Google's Antigravity. You prompt in plain language; the agent maps your request onto script runs. A two-step request becomes a two-command chain.

# You: "Make a 6-second clip of a cat drinking tea, then restyle it as anime."

# Step 1 - generate, capture the interaction id
./scripts/video/generate_video.py "A close-up of a cat drinking tea" \
  --duration 6 --output media/cat_tea.mp4

# Step 2 - restyle the exact clip from step 1
./scripts/video/generate_video.py "Transform the style to Japanese anime. Keep everything else the same." \
  --previous-interaction-id "<id from step 1>" --output media/cat_anime.mp4

The agent captures the interaction id from step one and passes it to step two, so the restyle edits the exact clip it just made. If you are deciding which agent to run this in, our Antigravity vs Claude Code comparison covers the trade-offs; the skill itself is agent-agnostic.

The Interpolation Discrepancy

One honest caveat. The skill exposes what it calls a video interpolation command, in which you pass two images, a start and an end, but the official Omni Flash API documentation lists interpolation and extension as not supported.

# The skill calls this "interpolation":
./scripts/video/generate_video.py "Move smoothly from the first frame to the last" \
  --image start.png --image end.png --output media/morph.mp4

So what actually happens? The two images are used as reference-to-video conditioning, not as true keyframes the model tweens between. The output is a plausible clip that starts and ends near those images, not a frame-accurate interpolation. Both the skill and the API docs are official; they simply describe the same capability with different words.

fofr's Subagent Music Video

The reason this skill matters is what it unlocks once an agent can call video generation on its own. The clearest demonstration came from fofr (@fofrAI), who published his own Omni API agent skill and then wired it into a subagent. He gave that subagent a hyperframes skill plus a batch of Omni Flash outputs, and it assembled a full music video on its own, with the soundtrack generated by Google's Lyria 3. No human in the timeline.

That is the endgame the gemini-omni-flash-api skill is built for: an agent that plans a sequence, generates each clip, edits it conversationally, and stitches the result, all through scripts it can actually run.

Companion Skill and MCP

The skill does not ship alone. First, gemini-interactions-api is a broader skill for the whole Interactions API, covering text, image, audio, and tool use, for agents that need more than video.

npx skills add google-gemini/gemini-skills --skill gemini-interactions-api

Second, Google ships a Gemini Docs MCP that gives a coding agent live access to the Gemini documentation, so the agent can look up current parameters instead of relying on training data. If you are assembling an agent toolkit, our best MCP servers for Claude Code roundup is a good starting point for pairing this skill with the right servers.

Limitations

The skill is honest scaffolding, but it inherits every constraint of the model underneath. Budget for these.

The regional block is the one that wastes the most time, because it does not throw an error. A blocked upload-edit just comes back fast and empty:

// Symptom of the regional upload-edit block
// (EEA, Switzerland, UK, some US states):
{
  "status": "completed",
  "usage": { "total_output_tokens": 0 },
  "steps": [ /* no model_output video block */ ]
}

FAQ

Glossary

Agent skill

A packaged set of instructions and scripts, described in a SKILL.md file, that a coding agent loads to gain a new capability.

gemini-omni-flash-preview

The model ID for Gemini Omni Flash, Google's natively multimodal video model, in public preview since June 30, 2026.

Interactions API

The Gemini endpoint (POST /v1beta/interactions) that Omni Flash runs through. It stores session state so you can chain edits.

previous_interaction_id

The field that continues a stored session, letting the model edit a prior clip without re-uploading it.

reference-to-video

An Omni Flash task type that conditions a new video on one or more reference images.

SynthID

The invisible, detectable watermark Google stamps on every Omni Flash output.

Verdict

The gemini-omni-flash-api skill is the right way to give a coding agent real Omni Flash video: it trades hallucinated API calls for tested scripts, and it slots into any skill-aware agent. It will not fix Omni's ceilings. The 720p cap, preview-grade schemas, per-second edit billing, and the regional upload lock are all still yours to manage.

Use it when:  your agent produces or iterates on short 720p clips in a pipeline.
Skip it when: you need 1080p/4K, guaranteed availability (it is preview),
              or you are in a region that blocks uploaded-video edits.
Always:       pin the SDK version and verify the "interpolation" flag
              is really reference-to-video before you promise a client.

Use it for agentic media pipelines where short, editable 720p clips are the deliverable. It is the difference between an agent that guesses at a video API and one that runs the same working command every time.

Sources

Official Google docs and repository:

• gemini-skills repository (the skill and its scripts)
• Gemini Omni Flash guide (interpolation and edit behavior)
• Omni Flash model page (720p, duration, aspect ratios)
• Interactions API reference
• Gemini API pricing (0.10 dollars per second of 720p video)
• Gemini API changelog (June 30, 2026 release)
• Google launch blog: Nano Banana 2 Lite and Gemini Omni Flash

Builders and community:

• fofr: published an Omni API agent skill
• fofr: subagent plus hyperframes auto-built a music video
• Philipp Schmid: Interactions API developer guide
• Logan Kilpatrick: Omni Flash launch and pricing