Which integrations are available for this server?

Provides vision capabilities via Google's Gemini API, enabling fast and cost-effective image description. Provides vision capabilities via NVIDIA's API, using models like meta/llama-3.2-11b-vision-instruct. Provides local/private vision inference using Ollama's models, such as llava, for offline image description. Provides general-purpose vision capabilities via OpenAI's API, enabling image description using GPT-4o and other models.

How do I use llm-vision-mcp?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@llm-vision-mcp Describe this image: ./images/sunset.jpg" That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

llm-vision-mcp

by ghbalf

Overview Schema Related Servers Score Discussions

TypeScript

Hybrid

llm-vision-mcp

An MCP server that gives vision capabilities to any LLM. It accepts images (file paths, URLs, or base64) and sends them to a vision-capable LLM, returning text descriptions that non-vision LLMs can use.

Providers

Provider	Default Model	Use Case
OpenAI	gpt-4o	General-purpose vision
Anthropic	claude-sonnet-4-latest	Detailed image analysis
Google	gemini-2.0-flash	Fast, cost-effective vision
Ollama	llava	Local/private inference
OpenAI-compatible	User-configured	DeepSeek, Qwen-VL, Together, etc.
Generic HTTP	N/A	Any API with custom request/response mapping

Related MCP server: VisionPower

Quick Start

npm install
npm run build

Option 1: CLI arguments (simplest)

node dist/index.js --provider openai --openai-api-key sk-...

Option 2: Environment variables

cp .env.example .env
# Edit .env with your API keys
node dist/index.js

Option 3: Config file (multi-provider)

cp config.example.json vision-config.json
# Edit vision-config.json
VISION_CONFIG_PATH=./vision-config.json node dist/index.js

MCP Client Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "vision": {
      "command": "node",
      "args": [
        "/absolute/path/to/llm-vision-mcp/dist/index.js",
        "--provider", "openai",
        "--openai-api-key", "sk-..."
      ]
    }
  }
}

Claude Code

Add to your .mcp.json:

{
  "mcpServers": {
    "vision": {
      "command": "node",
      "args": [
        "/absolute/path/to/llm-vision-mcp/dist/index.js",
        "--provider", "openai",
        "--openai-api-key", "sk-..."
      ]
    }
  }
}

Prompt: `vision_instructions`

The server registers an MCP prompt called vision_instructions that teaches the LLM when and how to use the describe_image tool. MCP clients that support prompts can inject this into the LLM's context so it automatically calls the tool whenever it encounters image paths, URLs, or base64 data — rather than guessing what an image contains.

Tool: `describe_image`

Sends an image to a vision LLM and returns a text description.

Parameters

Parameter	Required	Description
`image`	Yes	File path, URL, or base64-encoded image data
`prompt`	No	Custom instruction (default: "Describe this image in detail.")
`provider`	No	Override the default provider
`model`	No	Override the provider's default model

Image Input Formats

File path: /home/user/photo.png or ./images/chart.jpg
URL: https://example.com/image.png
Base64 data URL: data:image/png;base64,iVBOR...
Raw base64: Long base64 string (auto-detected)

Examples

"Describe this screenshot" + image: "/tmp/screenshot.png"
"Extract all text from this image" + image: "https://example.com/document.png"
"What data does this chart show?" + image: "data:image/png;base64,..."

Usage reporting

When the provider returns token counts, a second text content block is appended with Usage: <in> in / <out> out / <total> total tokens. Batch results (see describe_images) also include aggregated totalUsage.

Tool: `describe_images`

Describes multiple images in a single batched call. Each item may override the batch-level prompt, provider, and model. Results come back in input order. Per-provider concurrency limits are honored.

Parameters

Parameter	Required	Description
`items`	Yes	Array of 1–100 items, each with its own `image` and optional `prompt`/`provider`/`model`
`prompt`	No	Default prompt for items without their own
`provider`	No	Default provider for items without their own
`model`	No	Default model for items without their own
`concurrency`	No	Override the per-provider concurrency cap

Example call

{
  "items": [
    { "image": "/tmp/a.png" },
    { "image": "https://example.com/b.png", "prompt": "Extract text" }
  ],
  "prompt": "Describe this image in detail."
}

Sample result

{
  "results": [
    { "index": 0, "text": "A cat sitting on a desk.", "usage": { "inputTokens": 812, "outputTokens": 17, "totalTokens": 829 } },
    { "index": 1, "text": "Invoice header reading 'ACME Corp'." }
  ],
  "totalUsage": { "inputTokens": 812, "outputTokens": 17, "totalTokens": 829 }
}

Failed items appear with an error field instead of text; the batch itself does not fail.

Retry behavior

Transient errors — 429, 5xx, and network failures — are retried up to 3 times with exponential backoff. Configure via the top-level retry block (maxAttempts, baseDelayMs); per-provider retry overrides the global default.

Configuration

Configuration sources are loaded in this order (later overrides earlier):

.env file
Environment variables
CLI arguments
Config file (vision-config.json)
Per-request provider and model parameters

CLI Arguments

--provider <name>              Default provider
--openai-api-key <key>         OpenAI API key
--anthropic-api-key <key>      Anthropic API key
--google-api-key <key>         Google API key
--ollama-base-url <url>        Ollama URL (default: http://localhost:11434)
--ollama-model <model>         Ollama model (default: llava)
--model <model>                Default model for the default provider
--timeout <ms>                 Request timeout for the default provider
--ollama-timeout <ms>          Request timeout for Ollama (default: 120000)
--api-key <key>                API key for the default provider (generic)
--base-url <url>               Base URL for the default provider (generic)
--config <path>                Path to config file

Environment Variables

VISION_DEFAULT_PROVIDER=openai
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=AIza...
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=llava
VISION_TIMEOUT_MS=60000          # default provider timeout
OLLAMA_TIMEOUT_MS=300000         # bump for slow local models
VISION_CONFIG_PATH=./vision-config.json

# Preset providers — pick ONE vendor and replace the VISION_DEFAULT_PROVIDER
# value above; see "Preset Providers" section for the full list.
# Example (Moonshot):
#   VISION_DEFAULT_PROVIDER=moonshot
#   MOONSHOT_API_KEY=sk-...
#   # Optional: MOONSHOT_MODEL=kimi-k2.6, MOONSHOT_BASE_URL=https://api.moonshot.cn/v1/

Preset Providers

For 8 major OpenAI-compatible vision vendors, llm-vision-mcp ships with built-in preset defaults. Set VISION_DEFAULT_PROVIDER=<name> plus the vendor's standard API key env var — nothing else required. Optionally override the default model and base URL with <VENDOR>_MODEL / <VENDOR>_BASE_URL.

Preset name	Base URL	Default model	API key env var
`moonshot`	`https://api.moonshot.ai/v1/`	`kimi-k2.5`	`MOONSHOT_API_KEY`
`zai`	`https://api.z.ai/api/paas/v4/`	`glm-4.5v`	`ZAI_API_KEY`
`qwen`	`https://dashscope-intl.aliyuncs.com/compatible-mode/v1/`	`qwen3-vl-plus`	`DASHSCOPE_API_KEY`
`nvidia`	`https://integrate.api.nvidia.com/v1/`	`meta/llama-3.2-11b-vision-instruct`	`NVIDIA_API_KEY`
`groq`	`https://api.groq.com/openai/v1/`	`meta-llama/llama-4-scout-17b-16e-instruct`	`GROQ_API_KEY`
`together`	`https://api.together.xyz/v1/`	`meta-llama/Llama-Vision-Free`	`TOGETHER_API_KEY`
`deepinfra`	`https://api.deepinfra.com/v1/openai/`	`meta-llama/Llama-3.2-11B-Vision-Instruct`	`DEEPINFRA_API_KEY`
`xai`	`https://api.x.ai/v1/`	`grok-4.20-0309-non-reasoning`	`XAI_API_KEY`

Model strings use each vendor's exact casing (NVIDIA ships llama-3.2-11b-vision-instruct lowercase while DeepInfra ships Llama-3.2-11B-Vision-Instruct mixed case). Copy them verbatim — do not normalize.

Region notes:

zai — default baseUrl is the international endpoint (api.z.ai). Users in mainland China should override: ZAI_BASE_URL=https://open.bigmodel.cn/api/paas/v4/ (and use their bigmodel.cn-issued key as ZAI_API_KEY).
qwen — default baseUrl is the Singapore international endpoint. Users in mainland China should override: DASHSCOPE_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1/.

Quickstart with Moonshot:

export VISION_DEFAULT_PROVIDER=moonshot
export MOONSHOT_API_KEY=sk-...
llm-vision-mcp

The same pattern works for all 8 presets — swap moonshot and MOONSHOT_API_KEY for any other row of the table above.

MCP host config example (Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "vision": {
      "command": "node",
      "args": ["/absolute/path/to/llm-vision-mcp/dist/index.js"],
      "env": {
        "VISION_DEFAULT_PROVIDER": "moonshot",
        "MOONSHOT_API_KEY": "sk-..."
      }
    }
  }
}

Need multiple presets active at once, or pinned retry/concurrency settings per preset? See the Provider Cookbook below for copy-paste config-file snippets.

Config File

See config.example.json for a full example with all providers.

The config file supports ${ENV_VAR} interpolation — API keys can reference environment variables so they never appear in the file.

Provider Cookbook

Copy-paste JSON snippets for each preset vendor. Drop into your vision-config.json to pin settings, combine multiple providers, or override preset defaults. Keys stay in env vars via ${ENV_VAR} interpolation.

Moonshot (Kimi)

{
  "defaultProvider": "moonshot",
  "providers": {
    "moonshot": {
      "type": "openai-compatible",
      "baseUrl": "https://api.moonshot.ai/v1/",
      "apiKey": "${MOONSHOT_API_KEY}",
      "model": "kimi-k2.5"
    }
  }
}

Z.ai (Zhipu GLM) — international

{
  "defaultProvider": "zai",
  "providers": {
    "zai": {
      "type": "openai-compatible",
      "baseUrl": "https://api.z.ai/api/paas/v4/",
      "apiKey": "${ZAI_API_KEY}",
      "model": "glm-4.5v"
    }
  }
}

Z.ai (Zhipu GLM) — China region

Same vendor, different endpoint and key:

{
  "defaultProvider": "zai",
  "providers": {
    "zai": {
      "type": "openai-compatible",
      "baseUrl": "https://open.bigmodel.cn/api/paas/v4/",
      "apiKey": "${ZHIPUAI_API_KEY}",
      "model": "glm-4.5v"
    }
  }
}

Qwen (Alibaba DashScope) — international

{
  "defaultProvider": "qwen",
  "providers": {
    "qwen": {
      "type": "openai-compatible",
      "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/",
      "apiKey": "${DASHSCOPE_API_KEY}",
      "model": "qwen3-vl-plus"
    }
  }
}

NVIDIA NIM

{
  "defaultProvider": "nvidia",
  "providers": {
    "nvidia": {
      "type": "openai-compatible",
      "baseUrl": "https://integrate.api.nvidia.com/v1/",
      "apiKey": "${NVIDIA_API_KEY}",
      "model": "meta/llama-3.2-11b-vision-instruct"
    }
  }
}

Groq

{
  "defaultProvider": "groq",
  "providers": {
    "groq": {
      "type": "openai-compatible",
      "baseUrl": "https://api.groq.com/openai/v1/",
      "apiKey": "${GROQ_API_KEY}",
      "model": "meta-llama/llama-4-scout-17b-16e-instruct"
    }
  }
}

Together AI

{
  "defaultProvider": "together",
  "providers": {
    "together": {
      "type": "openai-compatible",
      "baseUrl": "https://api.together.xyz/v1/",
      "apiKey": "${TOGETHER_API_KEY}",
      "model": "meta-llama/Llama-Vision-Free"
    }
  }
}

DeepInfra

{
  "defaultProvider": "deepinfra",
  "providers": {
    "deepinfra": {
      "type": "openai-compatible",
      "baseUrl": "https://api.deepinfra.com/v1/openai/",
      "apiKey": "${DEEPINFRA_API_KEY}",
      "model": "meta-llama/Llama-3.2-11B-Vision-Instruct"
    }
  }
}

xAI (Grok)

{
  "defaultProvider": "xai",
  "providers": {
    "xai": {
      "type": "openai-compatible",
      "baseUrl": "https://api.x.ai/v1/",
      "apiKey": "${XAI_API_KEY}",
      "model": "grok-4.20-0309-non-reasoning"
    }
  }
}

Multiple providers simultaneously

{
  "defaultProvider": "openai",
  "providers": {
    "openai": {
      "apiKey": "${OPENAI_API_KEY}",
      "model": "gpt-4o"
    },
    "moonshot": {
      "type": "openai-compatible",
      "baseUrl": "https://api.moonshot.ai/v1/",
      "apiKey": "${MOONSHOT_API_KEY}",
      "model": "kimi-k2.5"
    },
    "zai": {
      "type": "openai-compatible",
      "baseUrl": "https://api.z.ai/api/paas/v4/",
      "apiKey": "${ZAI_API_KEY}",
      "model": "glm-4.5v"
    }
  }
}

Custom Providers

OpenAI-compatible (DeepSeek, Qwen-VL, etc.)

Most Chinese LLM providers expose an OpenAI-compatible API:

{
  "providers": {
    "deepseek": {
      "type": "openai-compatible",
      "baseUrl": "https://api.deepseek.com/v1",
      "apiKey": "${DEEPSEEK_API_KEY}",
      "model": "deepseek-vl2"
    }
  }
}

Generic HTTP (any API)

For APIs with non-standard request/response formats:

{
  "providers": {
    "custom": {
      "type": "generic-http",
      "url": "https://my-api.example.com/vision",
      "headers": { "Authorization": "Bearer ${API_KEY}" },
      "requestTemplate": {
        "image": "{{image}}",
        "prompt": "{{prompt}}",
        "type": "{{mimeType}}"
      },
      "imageFormat": "base64",
      "responsePath": "result.text"
    }
  }
}

Template placeholders: {{image}}, {{prompt}}, {{mimeType}}

imageFormat: "base64" (raw) or "data-url" (data:image/png;base64,...)

responsePath: Dot-notation path to extract the text from the JSON response (e.g., choices.0.message.content)

usagePath (optional): Dot-notation path to a numeric token total in the response. Reported as totalTokens.

Note: headers values are sent to the server literally — there is no ${ENV_VAR} expansion. Paste the bearer token directly, or launch the server from a wrapper that substitutes it.

Example: MiniMax vision (`MiniMax-M2.7`)

MiniMax's OpenAI-compatible /v1/chat/completions endpoint silently drops image_url content blocks, so vision requests must go to the Anthropic-compatible /anthropic/v1/messages endpoint with the image embedded as a plain-text data URL inside the content string (not as a content-part array).

{
  "providers": {
    "minimax": {
      "type": "generic-http",
      "url": "https://api.minimax.io/anthropic/v1/messages",
      "headers": { "Authorization": "Bearer YOUR_MINIMAX_API_KEY" },
      "requestTemplate": {
        "model": "MiniMax-M2.7",
        "max_tokens": 1024,
        "messages": [
          { "role": "user", "content": "{{image}}\n{{prompt}}" }
        ]
      },
      "imageFormat": "data-url",
      "responsePath": "content.0.text"
    }
  },
  "defaultProvider": "minimax"
}

Verified 2026-04 against MiniMax's international host (api.minimax.io). Chinese users swap the host for api.minimaxi.com. If MiniMax later ships a native image_url-style content part or adds a preset-class adapter in llm-vision-mcp, this generic-http config can be replaced by the simpler preset form.

Image Preprocessing

Images are automatically preprocessed before being sent to providers:

Format conversion: Unsupported formats (e.g., WEBP for providers that don't support it) are converted to PNG
Resizing: Images exceeding 2048x2048 are resized to fit (configurable)
Compression: Images exceeding 20MB are JPEG-compressed at decreasing quality levels

Preprocessing options can be customized in the config file:

{
  "preprocessing": {
    "maxWidth": 2048,
    "maxHeight": 2048,
    "maxFileSizeBytes": 20971520
  }
}

Development

npm test              # Run tests
npm run test:watch    # Watch mode
npm run build         # Compile TypeScript
npm run dev           # Watch mode compilation

License

ISC

This server cannot be installed

license - permissive license

quality - not tested

maintenance

How are these scores calculated?

Maintenance

–Maintainers

–Response time

2wRelease cycle

4Releases (12mo)

Commit activity

Resources

Need Help?

Related Servers

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Latest Blog Posts

Your AI Chatbot Just Exposed Your CEO's Salary to an Intern
By Om-Shree-0709 on July 2, 2026.
Agent Identity
MCP Security
OAuth Delegation
Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)
By Om-Shree-0709 on June 30, 2026.
Agentic Ai
Prompt Injection
WebAssembly
Lightport: Open-Sourcing Glama's AI Gateway
By punkpeye on April 27, 2026.
OpenAI
open source

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/ghbalf/llm-vision-mcp'

If you have feedback or need assistance with the MCP directory API, please join our Discord server

llm-vision-mcp

Providers

Quick Start

Option 1: CLI arguments (simplest)

Option 2: Environment variables

Option 3: Config file (multi-provider)

MCP Client Configuration

Claude Desktop

Claude Code

Prompt: vision_instructions

Tool: describe_image

Parameters

Image Input Formats

Examples

Usage reporting

Tool: describe_images

Parameters

Example call

Sample result

Retry behavior

Configuration

CLI Arguments

Environment Variables

Preset Providers

Config File

Provider Cookbook

Moonshot (Kimi)

Z.ai (Zhipu GLM) — international

Z.ai (Zhipu GLM) — China region

Qwen (Alibaba DashScope) — international

NVIDIA NIM

Groq

Together AI

DeepInfra

xAI (Grok)

Multiple providers simultaneously

Custom Providers

OpenAI-compatible (DeepSeek, Qwen-VL, etc.)

Generic HTTP (any API)

Example: MiniMax vision (MiniMax-M2.7)

Image Preprocessing

Development

License

Maintenance

Resources

Looking for Admin?

Latest Blog Posts

MCP directory API

Prompt: `vision_instructions`

Tool: `describe_image`

Tool: `describe_images`

Example: MiniMax vision (`MiniMax-M2.7`)