Gemini Media MCP: AI Planning, Storyboards & Costs
We’ve shipped a major update to our open-source Gemini Media MCP server. The server could already generate, compose, and edit images and video with Google’s Gemini, Veo, and Omni models. This release adds the production layer around those models: deciding what to make, previewing it before an expensive render, and understanding the cost before committing.
The result is a better workflow for AI agents, creative teams, and content operations. A prompt no longer has to jump directly into generation. It can become a ranked plan, a reviewable storyboard, a costed dry run, and then a finished asset.
AI Media Planning for Gemini and Veo
The new plan_generation tool reads a creative brief and returns ranked, ready-to-call production options. Each route includes the recommended tool and model, estimated cost, the parameters needed to run it, and an explanation of why other routes were ruled out.
For example, a request for a product-launch reel might produce a workflow that starts with inexpensive storyboard keyframes, moves through review, and then sends the approved shot list to generate_clip. A request for a landing-page hero may stay with image generation instead of proposing video that adds cost without serving the brief.
Planning is deterministic and rule-based, so it is instant and does not make another billable model call. It is also deployment-aware. If a model or workflow cannot run with the server’s current Gemini API or Vertex AI configuration, the planner will not recommend a path that is guaranteed to fail.
This matters when an agent has several technically valid ways to produce an asset. The best model is not always the most expensive model, and the best workflow is not always a single render. plan_generation makes those tradeoffs visible before production begins.
An AI Storyboard Generator for Veo Workflows
The new generate_storyboard tool renders a keyframe for every shot and composes the results into a real review artifact. It returns a bounded contact-sheet preview inline, while saving a full-resolution PNG and a self-contained HTML review page to disk.
Each panel can carry the information a production team needs to review the sequence: shot framing, prompt, camera direction, timing, and narrative order. The board can then feed directly into a multi-beat video workflow, with one aspect ratio carried from preview through final delivery.
That changes where feedback happens. Instead of discovering a sequencing, composition, or brand problem after several video renders, teams can review the visual argument as still frames first. Storyboards are faster and less expensive to regenerate, easier to share with stakeholders, and much clearer than approving a shot list from text alone.
For marketing teams, that makes AI video practical for campaign planning, product launches, explainers, social reels, and localized creative. The expensive step happens after the concept is visible and approved.
Gemini and Veo Cost Estimates Before Generation
All nine generation tools now support dry_run: true. A dry run validates the request and calculates the cost of the call that would actually execute without generating media.
That distinction is important. Estimates account for model substitutions, draft routing, video bridges, repeated extensions, storyboards, multi-beat clips, resolution constraints, and other details that change the real bill. A request that cannot run—because of an invalid option, inaccessible local file, disallowed GCS bucket, or deployment restriction—is refused during the quote instead of being assigned a misleading price.
Real runs report cost in both the response and sidecar manifest. When the server can inspect the delivered media, billing uses its measured duration rather than assuming the requested duration was returned. Every cost object also includes pricing provenance and the date of the embedded rate data.
This gives agents a useful control loop: plan several approaches, compare their costs, preview the chosen direction, and render only after the request passes validation.
Model Lifecycle and Production Hardening
This release also migrates discontinued Imagen endpoints and retired Gemini preview IDs to supported models while preserving compatibility for callers that still send legacy IDs. Substitutions are surfaced in responses, manifests, MCP notifications, and server logs rather than happening silently.
We also fixed Docker HTTP access, video bridge generation, storyboard font rendering, and a dry-run hang. Security and reliability improvements cover SSRF protection, local-file and GCS confinement, credential permissions, request timeouts, media limits, and stricter input validation. The release is backed by 1,196 passing tests, including wire-contract tests against the real Google SDK request shapes.
Available Now
Gemini Media MCP works with Claude Desktop, Claude Code, and any MCP-compatible platform. Developers and teams can run it from GitHub, Docker Hub, or PyPI.
It is also available on Adomo.ai, where teams can plan, generate, compose, and edit media through natural conversation without managing API keys or infrastructure.
For a closer look at fast generation and conversational video editing, read our earlier post on Gemini Omni MCP. For the complete implementation details behind this release, see pull request #9.
Thanks to the teams at Google and DeepMind for building these models.