gemini-vision-mcp-safe
A privacy-conscious MCP server that enables image analysis via Google Gemini with safety features like two-step confirmation and SSRF defenses.
README
gemini-vision-mcp-safe
English|简体中文
A minimal, privacy-conscious MCP server that lets an MCP client (Claude Code, Claude Desktop, etc.) ask Google Gemini to look at an image — local file or URL — and return a description, text extraction, comparison, and so on.
The "safe" in the name is a design goal:
- Two-step handshake before any image leaves the machine. The first call
returns a Chinese confirmation prompt; the second call (with
confirm_send_to_gemini=true) is the only one that actually talks to Gemini. This stops a model from silently uploading user files to Google. - SSRF defenses for URL inputs: protocol allowlist, manual redirect handling, per-hop DNS check against private/loopback ranges, HTTPS→HTTP downgrade refused.
- Magic-byte sniffing instead of trusting file extensions or
Content-Type. - Configurable size cap with both
Content-Lengthpre-check and a hard streaming limit; remote bodies stream to a temp file that is removed infinally. - Proxy aware via
HTTPS_PROXY/HTTP_PROXY. Useful where Google APIs are not directly reachable (e.g. mainland China through clash/mihomo). The proxy URL is redacted in logs. - API key in
.env, not in the MCP config. The repo's.gitignoreexcludes.envso the key never ends up in git.
Result text and error messages are in Chinese.
Tools
analyze_image_with_gemini
Send one local image or one HTTP/HTTPS image URL to Gemini.
| Parameter | Required | Description |
|---|---|---|
image_source |
yes | Local path (C:/path/to.png) or URL (https://…). |
prompt |
no | What to ask Gemini. Defaults to a Chinese "describe this image" prompt. |
model |
no | Override the model for this call (e.g. gemini-2.5-flash). Falls back to GEMINI_VISION_MODEL. |
confirm_send_to_gemini |
no | Must be true to actually send. Defaults to false. |
analyze_images_batch
Send 2–5 images in a single Gemini call (good for "compare these screenshots" or multi-page documents).
| Parameter | Required | Description |
|---|---|---|
image_sources |
yes | Array of 2–5 paths or URLs. Mixed is fine. |
prompt |
no | What to ask Gemini across all images. |
model |
no | Same as above. |
confirm_send_to_gemini |
no | Same handshake. |
If one image fails to load, the error message tells you which one
(第 N 张: …).
Install
git clone https://github.com/nianshou555qiansui/gemini-vision-mcp-safe.git
cd gemini-vision-mcp-safe
npm install
npm run build
cp .env.example .env # then edit .env, paste your Gemini API key
Get a key at https://aistudio.google.com/apikey.
Wire it up
Claude Code
~/.claude.json (or claude_desktop_config.json for Claude Desktop):
{
"mcpServers": {
"gemini-vision-safe": {
"type": "stdio",
"command": "node",
"args": [
"--env-file=/absolute/path/to/gemini-vision-mcp-safe/.env",
"/absolute/path/to/gemini-vision-mcp-safe/dist/index.js"
],
"env": {
"HTTPS_PROXY": "http://127.0.0.1:7890",
"HTTP_PROXY": "http://127.0.0.1:7890"
}
}
}
}
--env-file requires Node ≥ 20.6. The env block in MCP config is for
non-secret settings (proxy address); the API key lives in .env so it
never ends up in version control or shared configs.
On Windows where the launcher needs a shell, use cmd /c node … instead of
node ….
Configuration
.env keys (see .env.example):
| Key | Default | Notes |
|---|---|---|
GEMINI_API_KEY |
(required) | Your key from Google AI Studio. |
GEMINI_VISION_MODEL |
gemini-2.5-flash |
Default model. Per-call model arg overrides this. |
GEMINI_VISION_MAX_IMAGE_MB |
10 |
Hard cap. Remote images that exceed this via Content-Length are rejected before download; the streaming reader also enforces it. |
GEMINI_VISION_REQUEST_TIMEOUT_MS |
20000 |
Per-hop URL fetch timeout. |
GEMINI_VISION_GEMINI_TIMEOUT_MS |
60000 |
SDK-level timeout for the Gemini call. |
GEMINI_VISION_ALLOW_URL |
true |
Set false to refuse URL inputs entirely. |
GEMINI_VISION_ALLOW_LOCAL_FILE |
true |
Set false to refuse local file inputs. |
GEMINI_VISION_BLOCK_LOCAL_URLS |
true |
Set false to disable SSRF/private-IP blocking (not recommended). |
HTTPS_PROXY / HTTP_PROXY |
unset | Used by both fetch and the Gemini SDK via undici's global dispatcher. |
Privacy notes
- A local file path stays local; only the bytes of the file you confirm travel to Gemini.
- A URL is fetched from your machine first, then forwarded to Gemini — the original host sees your IP (or your proxy's), but never sees Google. Conversely Google never sees the original host.
- The repo never contains an API key. Verify before committing:
git ls-files | grep -F .envshould print nothing.
Caveats
- DNS rebinding / TOCTOU is not mitigated: the SSRF check uses the OS resolver, but the actual TCP connect resolves again. Acceptable for local use; not safe to expose this MCP as a public service.
- The OS resolver does not go through
HTTPS_PROXY. If your local DNS is unreliable, prefer URLs whose hostnames you've already pre-resolved or run a trusted DNS upstream. - Per-region Gemini availability is Google's decision. The error mapper will tell you when it sees 403 / 429 / 503.
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.