CXO AGI
Back to Blog Gemini Omni MCP: fast video generation and editing

Gemini Omni MCP: fast video generation and editing

CxO AGI Team •

We’ve shipped v0.1.9 of our open-source Gemini Media MCP server, and it’s the biggest release yet. This one adds Google’s gemini-omni-flash-preview model alongside the existing Veo tools. For AI agents that make video, it adds two things: fast generation and conversational editing.

Fast Video, Then Conversational Editing

Omni generates 720p video significantly faster than Veo, fast enough to iterate in real time. What happens after the first generation matters more.

With conversational editing, you pass the interaction_id from a previous result and describe only the change, such as “make the sky stormy,” “slow the pan,” or “add a neon sign in the window.” The server holds the video context, so there are no re-uploads and no starting over. Chain edits turn after turn until the shot is exactly right, the same way you’d direct a human editor.

generate_video_omni → interaction_id
edit_video(interaction_id, "make the sky stormy") → new interaction_id
edit_video(interaction_id, "add rain on the lens") → ...

That’s a better loop for any AI agent producing video.

Draft on Omni, Finalize on Veo

The new draft flag lets you run the same generate_video call on Omni for a fast, cheap preview, then flip it back to render the final on Veo at full quality (1080p/4K, seeds, frame control). And the new animatic flag on generate_clip renders every beat of a multi-beat reel through Omni first, so you can preview the whole storyboard as a 720p animatic before committing to full Veo renders.

You draft cheaply and render at full quality only once, so nothing is wasted.

What’s New in v0.1.9

New tools

  • generate_video_omni generates fast 720p video from a text prompt, with optional reference images (multi-image supported) or an input video to transform. Returns an interaction_id for follow-up edits.
  • edit_video handles conversational editing. Pass the interaction_id and describe only the change.
  • loop_extend extends a Veo video by ~7s, up to 20 times in one call, chaining each result into the next.

New controls

  • Video: resolution (720p/1080p/4K on Veo), person_generation, per-call timeout_seconds on Omni tools, output_gcs_uri for Omni on Vertex.
  • Image: aspect_ratio and person_generation across all Gemini image models.
  • Responses and sidecar manifests now include a warnings field whenever a request couldn’t be fully honored, so parameters are never dropped silently.

Model refresh

  • Added GA image models: gemini-3-pro-image, gemini-3.1-flash-image, and gemini-3.1-flash-lite-image (the cheapest option yet).
  • Veo model IDs now translate automatically per backend (-001 on Vertex ↔ -preview on the Gemini API), which fixes 404s on Gemini-API-only deployments.

Built to Be Trusted in Production

An MCP server has to behave predictably under load, not only expose capabilities. This release includes a hardening pass:

  • Fail-loud everywhere. Unfetchable inputs and unsupported model/mode combinations now error clearly instead of silently changing your output.
  • SSRF fix. HTTP redirects are re-validated hop-by-hop against the private-IP guard.
  • Truthful manifests. Real durations, real audio state, size caps on base64 inputs, consistent GCS bucket allowlisting, and accurate MIME detection for Omni inputs.
  • Setup wizard now validates credentials with a real, time-bounded API call.

It works with both the Gemini API and Vertex AI.

Available Live on Adomo.ai

You don’t need to install anything to use Omni. It’s live right now on Adomo.ai, our enterprise AI platform, with the full Gemini Media MCP built in: fast Omni drafts, conversational editing, and cinematic Veo finals, all through natural conversation.

Sign up at Adomo.ai and start generating and editing video in minutes, with no setup, API keys, or infrastructure to manage.

Why We Open-Sourced It

We originally built Gemini Media MCP for Adomo AI’s enterprise platform, then open-sourced it because every AI agent should be able to create professional visual content. With Omni it covers fast previews, conversational editing, cinematic finals, the latest image models, and production-grade reliability, all through natural conversation.

Try it live on Adomo.ai, no setup required.

For developers and teams who want to run it themselves:

  • GitHub, full source code and documentation
  • Docker Hub, pull and run in seconds
  • PyPI, install via pip

It works with Claude Desktop, Claude Code, and any MCP-compatible platform.

A Few Notes

  • Omni output is fixed at 720p / 24fps, 3–10s, so draft on Omni and finalize on Veo for high-res output.
  • Conversational editing on Vertex is preview-status upstream and may require allowlisting; run one smoke edit before relying on it.
  • Requires google-genai >= 2.10.0.
  • ⚠️ Breaking changes: imagen-3.0-generate-002 was removed (retired by Google). Unsupported aspect_ratio values now error instead of coercing to 16:9 (supported: 16:9, 9:16), and unfetchable image_uri/last_frame_uri/reference URIs now error instead of falling back to text-to-video.
  • ⏰ Imagen 4.x shuts down 2026-08-17 (Google). Results now carry a runtime deprecation warning, so migrate to the gemini-3.x image family.

Thanks to the teams at Google and DeepMind for building these models.