Gemini媒体生成MCP服务器

gemini-media-mcp

An MCP server for Google's generative media models: images (Nano Banana 2.1 / Pro), video (Veo 3.1 and Gemini Omni Flash), speech (Gemini 3.8 TTS) and music (Lyria 3.5). It ships as a single Go binary, speaks stdio and Streamable HTTP (MCP 2026-07-28), works with the Gemini API or Vertex AI, and comes with agent skills that teach agents the whole workflow, plus plugin packaging for Claude Code, Codex, Gemini CLI, VS Code/Copilot, Cursor and more.

https://github.com/user-attachments/assets/f46970b3-82a6-4364-a40a-922103dbbc70

An agent using the server one tool at a time: images and edits, a video clip and its claymation restyle, a 768 px portrait tiled up to 8K, and an extended shot. The footage, the narration and the music all came from the server.

• Current models, updated without a release. A built-in catalog records IDs, aliases, lifecycle, parameters and prices. Retired models redirect to their replacement, and new model IDs work before the catalog knows them. You can override or extend the catalog with a hot-reloaded YAML file.
• Cost-aware. Every result reports its estimated cost, and estimate_cost compares options before you spend. Spend is recorded in a ledger, capped by session, daily and monthly budgets, and calls above a threshold need explicit approval.
• Agent-friendly. Each tool matches a workflow, and errors come back as [kind] message + Hint. Image results include inline previews, and outputs can be chained by URI. Video runs as async jobs with long-polling and progress notifications.
• Skills included. Six agent skills cover each medium, multi-asset productions and upscaling past 4K. The upscale skill re-renders a photo tile by tile and stitches the tiles into an 8K image.
• Robust. Backend and credentials are detected the way the Google SDKs do it. Vertex locations are chosen per model, retries and timeouts are built in, files are written atomically with provenance, and HTTP mode ships with security defaults.

Upgrading from v0? The tools changed. See the migration table. The review also covers what was broken, why, and the design of v1.

Quick start

1. Get credentials. Either:

   • an API key from Google AI Studio (GEMINI_API_KEY), or
   • a Google Cloud project with Vertex AI enabled (GOOGLE_CLOUD_PROJECT plus gcloud auth application-default login).

2. Install the binary. On macOS or Linux:

   复制

   mkdir -p ~/.local/bin   # must be on your PATH
   os=$(uname -s | tr '[:upper:]' '[:lower:]'); arch=$(uname -m | sed 's/x86_64/x64/; s/aarch64/arm64/')
   curl -fsSL "https://github.com/mordor-forge/gemini-media-mcp/releases/latest/download/$os.$arch.gemini-media-mcp.tar.gz" \
     | tar -xz -C ~/.local/bin gemini-media-mcp
   gemini-media-mcp version
   

   On Windows, download the windows zip from the latest release and put gemini-media-mcp.exe on your PATH. With Go 1.26+: go install github.com/mordor-forge/gemini-media-mcp/cmd/gemini-media-mcp@latest.

   The Gemini CLI extension, the Claude Desktop bundle and the Docker image include the binary, so they skip this step.

3. Add the server to your agent. Every config below runs gemini-media-mcp from your PATH.

   Claude Code (plugin: server plus skills), from the shared mordor-forge marketplace:

   复制

   /plugin marketplace add mordor-forge/claude-plugins
   /plugin install gemini-media@mordor-forge
   

   If you added mordor-forge from this repository before, switch with /plugin marketplace remove mordor-forge first. Your installed plugin keeps working either way. or just the server: claude mcp add gemini-media -e GEMINI_API_KEY=... -- gemini-media-mcp

   Codex (~/.codex/config.toml):

   复制

   [mcp_servers.gemini-media]
   command = "gemini-media-mcp"
   env_vars = ["GEMINI_API_KEY", "GOOGLE_API_KEY", "GOOGLE_CLOUD_PROJECT", "GOOGLE_CLOUD_LOCATION"]
   tool_timeout_sec = 300   # 4K images and video polls can exceed the 60 s default
   

   Gemini CLI (paid API keys, Vertex AI and Gemini Code Assist only, since June 2026): gemini extensions install https://github.com/mordor-forge/gemini-media-mcp

   Antigravity CLI: add the server to ~/.gemini/config/mcp_config.json, as shown in the install snippets.

   Claude Desktop: download gemini-media-mcp-<version>.mcpb from the latest release and open it.

   VS Code / Copilot: code --add-mcp '{"name":"gemini-media","command":"gemini-media-mcp"}'

   Cursor, Windsurf and other mcpServers-style clients (Zed, OpenCode and Goose use their own formats, see below):

   复制

   { "mcpServers": { "gemini-media": { "command": "gemini-media-mcp", "env": { "GEMINI_API_KEY": "..." } } } }
   

   If a desktop app reports that gemini-media-mcp was not found, it doesn't see your shell's PATH: use the full path from command -v gemini-media-mcp. Every client, plus Docker and the skills installer, is covered in packaging/INSTALL-SNIPPETS.md.

4. If your agent doesn't forward environment variables (plugins often don't), store the key once:

   复制

   echo "$GEMINI_API_KEY" | gemini-media-mcp configure --api-key-stdin
   gemini-media-mcp doctor   # checks credentials, backend and model availability
   

To run it in Docker instead: docker run -i --rm --user "$(id -u):$(id -g)" -e GEMINI_API_KEY -v "$PWD/media:/output" -v gemini-media-state:/state ghcr.io/mordor-forge/gemini-media-mcp (--user makes the generated files yours on Linux; the named /state volume keeps spend accounting and video jobs between runs).

Tools

Every result includes the saved file's path, a gemini-media://files/<name> URI and its cost. You can pass the URI as an input to another tool. Clients that can't read the server's disk (for example over HTTP) can fetch the file with resources/read, for files up to 64 MB (larger upscales stay on the server's disk).

Skills

The skills/ directory contains Agent Skills that teach agents the full workflow for each media type: intent, prompt craft, model choice, cost checks, review and iteration. They cover interactive use as well as unattended runs.

Plugin installs (Claude Code, Codex, Gemini CLI, VS Code) include the skills. To install them in any Agent Skills–compatible agent, run npx skills add mordor-forge/gemini-media-mcp, or copy the folders into .agents/skills/ or ~/.claude/skills/.

Models

Use an alias or a full model ID. Run gemini-media-mcp models for the built-in catalog with prices, or call list_models, which also applies your override file and, with live: true, checks what your key can use.

Veo on the Gemini API. Google is shutting down the Veo 3.1 previews (and the gemini-omni-flash-preview model) on the Gemini API on 2026-10-22. Until then lite, fast and standard still work there with a deprecation warning; after it the server uses omni instead and says so. Veo 3.1 stays available on Vertex AI, now called Gemini Enterprise Agent Platform (veo-3.1-*-generate-001): to keep using it, run the server against a Google Cloud project: GOOGLE_CLOUD_PROJECT plus gcloud auth application-default login, and GEMINI_MEDIA_BACKEND=vertex if an API key is also set (see Configuration).

Updating or adding a model without waiting for a release. Create a YAML file and point GEMINI_MEDIA_CATALOG at it. The server merges it with the built-in catalog by id and reloads it automatically:

defaults:
  video: fast
models:
  - id: veo-3.1-fast-generate-preview
    pricing: { perSecond: { 720p: 0.10, 1080p: 0.12, 4k: 0.30 } }
  - id: veo-4.0-generate-preview        # a brand-new model
    aliases: [veo4]
    pricing: { perSecond: { 720p: 0.50 } }

See internal/catalog/models.yaml for the full schema.

Configuration

Settings are layered: built-in defaults < config file < environment < flags. The config file lives at ~/.config/gemini-media-mcp/config.yaml on Linux, ~/Library/Application Support/gemini-media-mcp/config.yaml on macOS, %AppData%\gemini-media-mcp\config.yaml on Windows, or wherever GEMINI_MEDIA_CONFIG points. Unknown keys produce warnings instead of errors, and get_config shows where each setting came from.

How the backend is chosen:

1. An explicit backend setting wins.
2. Otherwise the SDK switches GOOGLE_GENAI_USE_ENTERPRISE and GOOGLE_GENAI_USE_VERTEXAI decide.
3. Otherwise an API key selects the Gemini API, even if GOOGLE_CLOUD_PROJECT is set in your shell for other tools.
4. Otherwise a project selects Vertex AI.

On Vertex, an API key without a project uses express mode, which does not support video.

Spend and budgets

Google doesn't return costs, so the server computes them from its price table and the token usage the API reports. Every call records an estimate before it runs and the reconciled cost afterwards. Failed and safety-blocked generations count as $0. A response that used tokens but returned no media (for example MAX_TOKENS) is charged for those tokens.

• Results include cost {estimatedUsd, usd, basis}. get_usage and gemini-media-mcp usage summarize spend. The ledger is a plain JSONL file in the state directory.
• Budgets are enforced before each call, across concurrent calls and across several server processes sharing a state directory (in-flight calls hold their estimate under a file lock).
• With GEMINI_MEDIA_CONFIRM_ABOVE_USD=1, an expensive call (for example a $3.20 Veo clip) comes back as a [confirmation] error. The agent asks you, then retries with approvedCostUsd.
• These figures are estimates. For a hard stop, also set a spend cap in AI Studio or a budget in Cloud Billing.

HTTP mode

GEMINI_API_KEY=... gemini-media-mcp serve --transport http                 # http://127.0.0.1:8765/mcp
GEMINI_MEDIA_HTTP_TOKEN=secret gemini-media-mcp serve --transport http --http-addr 0.0.0.0:8765
claude mcp add --transport http gemini-media http://127.0.0.1:8765/mcp --header "Authorization: Bearer secret"

• HTTP mode is stateless Streamable HTTP, which MCP 2026-07-28 requires; older clients still work.
• The server protects against DNS rebinding and cross-origin requests.
• It refuses to listen on a non-loopback address without a token.
• It only reads input files from the output directory or GEMINI_MEDIA_INPUT_DIRS (unless GEMINI_MEDIA_ALLOW_ANY_INPUT_PATH=true).
• GET /healthz reports status.

Development

Paid live E2E tests are run locally with your own API key. CI runs the free checks and compiles/vets the E2E tests, but never executes live generation, including after merges to main or on manual workflow runs.

go build ./...
go test -race ./...
golangci-lint run
python3 scripts/validate-skills.py skills
GEMINI_MEDIA_E2E=1 GEMINI_API_KEY=... go test -tags=e2e ./internal/media/ -run E2E -v -timeout 30m   # live, costs cents
# add GEMINI_MEDIA_E2E_VIDEO=1 for the video test (~$0.20), GEMINI_MEDIA_E2E_OMNI=1 for the Omni clip + edit (~$0.30),
# GEMINI_MEDIA_E2E_UPSCALE=1 for a small tiled upscale (~$0.20),
# GEMINI_MEDIA_E2E_OUTPUT_DIR=./e2e-out to keep the files

• AGENTS.md: guide for coding agents and contributors.
• docs/architecture-review.md: architecture and design decisions.
• internal/catalog/models.yaml: model catalog. When Google changes models, edit this file, not the code.

License

Apache-2.0