Agent Tools for Genmedia

Agent Tools for Genmedia is the entry point for the agent-facing tooling that lets AI agents and applications drive Google Cloud’s generative media APIs (Gemini Image, Gemini TTS, Veo, Chirp, Lyria, and audio/video compositing with AVTool).

It corresponds to the experiments/agent_tools/ directory in the repository, which is the intended home for a consolidated set of genmedia agent tooling: the existing MCP Servers and Agent Skills, and — over time — Agent Plugins, gathered together in one place.

Scope today

Right now this directory holds one artifact: a cross-server smoke test for the media-generation MCP servers (described below). It does not move or replace the existing MCP Servers for Genmedia — those stay where they are and continue to be the way you run the servers. Bringing Agent Skills and Agent Plugins into this directory is future work, and the exact approach is still being decided; nothing here should be read as a committed migration plan or timeline.

MCP Servers for Genmedia

The Model Context Protocol (MCP) servers are the shipped, production way to give an agent access to Google Cloud genmedia. Each server can be run independently, and they cover image, video, speech, music, and audio/video compositing.

➡️ See the MCP Servers for Genmedia overview for the full list of servers, installation, configuration, and per-server documentation.

Agent Skills and Agent Plugins

The repository already ships a set of genmedia Agent Skills — markdown expertise that teaches an agent how to use the MCP tools for complex workflows (for example genmedia-producer, genmedia-video-editor, and genmedia-audio-engineer). These are documented today under the MCP GenMedia section; see Agent Skills.

Consolidating these skills — and packaging genmedia tooling as Agent Plugins — into experiments/agent_tools/ is a direction under exploration, not a decided plan or a shipped layout. This page will grow to point at them as that work lands.

The generate-and-verify smoke test

The first artifact to live in experiments/agent_tools/ is smoke_generate_and_verify.sh: a consolidated generate-and-verify smoke test across the mcp-genmedia-go servers. For each server it fires one realistic media-generation tools/call (via the external mcptools CLI) and then verifies that a real media artifact was produced — a non-empty local file, or a GCS object that gcloud storage ls can see.

This is intentionally stronger than each server’s existing verify.sh, which only does a go build + tools/list liveness check and never produces media.

What it covers

Server	Tool called	Notes
mcp-gemini-go	gemini_image_generation
mcp-nanobanana-go	nanobanana_image_generation
mcp-veo-go	veo_t2v	Video generation; can take minutes. Veo writes to GCS, so GCS mode is recommended for this server.
mcp-lyria-go	lyria_generate_music
mcp-chirp3-go	chirp_tts	Local output only (no GCS output parameter).
mcp-omni-go	omni_video_generation	Video generation; can take minutes.
mcp-avtool-go	ffmpeg_convert_audio_wav_to_mp3	Chained off the chirp output (converts chirp’s .wav to .mp3); SKIPped if ffmpeg is unavailable or no chirp .wav was produced.

mcp-common is a shared library, not a server, and is skipped. mcp-imagen-go is intentionally not covered: Imagen models were shut down across Google (including Vertex AI) on 2026-08-17 and return HTTP 404.

Requirements

• mcptools: go install github.com/f/mcptools/cmd/mcptools@latest (ensure $(go env GOPATH)/bin is on your PATH).
• A go toolchain (each server is built before it is called) and jq.
• For GCS mode: gcloud with application-default credentials.
• GOOGLE_CLOUD_PROJECT must be set (required by every server).
• ffmpeg — only needed by mcp-avtool-go; if it is absent, avtool is SKIPped (not failed) and every other server still runs.

Usage

# Required
export GOOGLE_CLOUD_PROJECT=your-project

# Optional: write generated media to GCS and verify with `gcloud storage ls`.
# If unset, media is written locally under ./smoke_output/ instead.
export GENMEDIA_BUCKET=gs://your-bucket/some/prefix

# Run all servers
./smoke_generate_and_verify.sh

# Run a subset
./smoke_generate_and_verify.sh veo lyria chirp

# Increase the per-call timeout (default 600s) for slow video models
SMOKE_CALL_TIMEOUT=900 ./smoke_generate_and_verify.sh

A per-server summary table is printed at the end, and the exit code is non-zero if any server fails to produce verified media (a SKIP does not fail the run). Verification never trusts the JSON-RPC response alone — it confirms the artifact actually exists (listing the GCS destination prefix, or checking for a non-empty local file), which matters because some servers, notably mcp-veo-go, return a resource_link content type that the mcptools CLI cannot render even though the media is real.

For the complete, authoritative details, see the experiments/agent_tools/README.md in the repository.