MCP Servers for Google Cloud Genmedia APIs

This repository provides Model Context Protocol (MCP) servers that enable AI agents and applications to easily integrate and leverage the full breadth and depth of Google Cloud’s powerful generative media APIs (Gemini Image, Gemini TTS, Veo, Chirp, Imagen, Lyria) and advanced audio/video compositing capabilities (AVTool).

Each server can be enabled and run separately, allowing flexibility for environments that don’t require all capabilities.

Generative Media & Compositing Capabilities

Nano Banana: Gemini 3.1 Flash Image & Gemini 3 Pro Image & Gemini 2.5 Flash Image - for image generation and editing
Veo 3 & 3.1 - for video creation
Gemini TTS & Chirp 3 HD - for speech synthesis
Lyria - for music generation
Imagen 3 & 4 - for image generation and editing
AVTool - for audio/video compositing and manipulation

🛠️ Agent Skills

We provide a set of expert Agent Skills that provide AI agents (like Gemini CLI or Antigravity) with deep domain knowledge on how to use these MCP tools effectively for complex workflows.

genmedia-producer: Orchestrates multi-step workflows like podcast creation and storyboarding.
genmedia-video-editor: Expertise in FFmpeg composition, image overlays, and GIF generation.
genmedia-audio-engineer: Specialist in high-fidelity TTS synthesis and multi-track mixing.
genmedia-image-artist: Expert in visual generation, prompt optimization, and collaborative refinement.
genmedia-voice-director: Expert in casting, directing, and generating expressive text-to-speech using Gemini TTS.

See the Agent Skills directory for more information on how to link or install these skills.

Installation

Install MCP Servers: For detailed installation instructions, including an easy-to-use installer script, please refer to the Go Implementations README.

Easy Installation (Pre-compiled Binaries)

For the fastest setup without needing the Go toolchain installed, you can use our online installer script. This script automatically detects your operating system and architecture, downloads the latest pre-compiled binaries from GitHub Releases, and places them in your ~/.local/bin directory.

Run the following command in your terminal:

curl -sL https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-creative-studio/main/experiments/mcp-genmedia/mcp-genmedia-go/install-online.sh | bash

Note: Ensure ~/.local/bin is added to your system PATH.

Running the Servers

The MCP servers can be run using different transport protocols. The default is stdio.

To start a server in Streamable HTTP mode, use the --transport http flag:

mcp-imagen-go --transport http

Configuration

The servers are configured primarily through environment variables. Key variables include:

PROJECT_ID: Your Google Cloud project ID.
LOCATION: The Google Cloud region for the APIs (e.g., us-central1).
PORT: The port for the HTTP server (e.g., 8080).
GENMEDIA_BUCKET: The Google Cloud Storage bucket for media assets.

Available MCP Servers and Capabilities

Gemini Image Generate and edit images from text prompts.
Gemini TTS Synthesize high-quality audio from text.
Veo: Create videos from text or images.
Lyria: Generate music from text prompts.
Chirp 3 HD: Synthesize high-quality audio from text.
Imagen: Generate and edit images from text prompts.
AVTool: Perform audio/video compositing and manipulation (e.g., combining, concatenating, format conversion).

For a detailed list of tools provided by each server, refer to the Go Implementations README.

Naming Outputs: `output_filename`

Every media-generating tool accepts a single optional output_filename (string) with identical name and semantics across servers, so a client can predict the exact output name from the tool and request. The same computed name is applied consistently to the inline response, the local output_directory, and GCS.

Precedence: output_filename always wins over a server’s legacy naming parameter; if neither is set, the server’s default scheme is used. When output_filename is unset, behavior is unchanged (no regression).
Extension forced to the true type: you provide the file stem and the extension is set from the model’s actual output MIME type (hero.jpeg on a PNG response → hero.png; a missing extension is added). AVTool is exempt — for ffmpeg the extension selects the output container, so AVTool keeps the extension you provide.
Multiple outputs → _1..n suffix: one output is <stem>.<ext>; n > 1 outputs are <stem>_1.<ext> … <stem>_n.<ext> (1-based, no zero-padding, in generation order).
Collisions overwrite with a warning (re-running the same name is idempotent), and path traversal is sanitized to a single safe name.
Imagen & Veo let the Vertex API name objects (sample_*) and then rename them to output_filename; on success no sample_* originals remain.

Deprecated legacy aliases (still accepted; prefer output_filename, no removal planned): AVTool output_file_name, Lyria file_name, and the Chirp 3 / Gemini TTS output_filename_prefix (a prefix only — output_filename gives the full name).

Resource Links for GCS Outputs

When a tool writes an artifact to Google Cloud Storage, it additionally returns one MCP resource_link content item per GCS artifact, so an MCP client can address the output as a first-class resource rather than parsing it out of the text summary.

GCS sink only. A resource_link is emitted only when the artifact was uploaded to GCS (gcs_bucket_uri or a server’s bucket fallback). Inline-only and local-only responses are unchanged — no link is added.
One link per artifact, appended after the existing content in generation order (so n GCS artifacts produce n links, aligned 1..n).
uri is the gs:// URI — the durable identity of the object, not the expiring V4-signed HTTPS URL. The signed URL (when generated) still appears in the text summary; the resource_link gives clients a stable handle.
name is the object’s name, mimeType is the artifact’s true output MIME type, and description is a 1-based human label (e.g. nanobanana output 1 of 1).
Text output is unchanged (back-compat). The resource_link items are appended; the existing text content is byte-for-byte identical to before this feature.

In scope: Gemini Image, Imagen (text-to-image and edit), Veo, Lyria, Nanobanana. Out of scope: Gemini TTS, Chirp 3, and AVTool do not emit resource_link items.

Authentication

The servers use Google’s Application Default Credentials (ADC). Ensure you have authenticated by one of the following methods:

Set up ADC: gcloud auth application-default login
Set the GOOGLE_APPLICATION_CREDENTIALS environment variable to the path of your service account key file.

You may also need to grant your user or service account access to the Google Cloud Storage bucket:

gcloud storage buckets add-iam-policy-binding gs://BUCKET_NAME \
  --member=user:user@email.com \
  --role=roles/storage.objectUser

Client Configurations

The MCP servers can be used with various clients and hosts. A sample MCP configuration JSON can be found at genmedia-config.json.

This repository provides AI application samples for:

Development and Contribution

For those interested in extending the existing servers or creating new ones, the mcp-genmedia-go directory contains a more detailed README.md with information on the architecture and development process. Please refer to the mcp-genmedia-go/README.md for more information.

License

Apache 2.0

Disclaimer

This is not an officially supported Google product.