Cove Sensory MCP
A local MCP server sends images, video, audio and music to a multimodal model and returns a description to a text-only agent.
- 120 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Images, video and audio go to the chosen provider under your API key and are billed by it.
- The server reads local files from the directories you allow.
Install
- Works in
Manual install
git clone https://github.com/moonlin1213/cove-sensory-mcp.gitThe repository ships as source code. After cloning, run `uv sync --locked` and `uv run cove-sensory-mcp doctor` as in the README.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Images, video and audio go to the chosen provider under your API key and are billed by it.
- The server reads local files from the directories you allow.
This is third-party code. Review the repository files before installing.
What it does
An agent without sight or hearing gets seven tools: `sense_image`, `sense_video`, `sense_audio`, `sense_music` and three service tools, `sensory_status`, `sensory_setup_guide` and `sensory_self_test`. The server passes an explicitly named file to a configured provider and returns a structured observation. Memory, personality, playback and continuous monitoring are outside its scope. Gemini or MiniMax-M3 serve as eyes, and Gemini or an OpenAI-compatible provider with the `audio_url_data_uri` mode, such as Qwen3-Omni through SiliconFlow, as ears. Keys live in the OS credential store. Key entry goes through `cove-sensory-mcp configure`, never through chat.
Who it is for. Users of text-only models with no built-in vision who need a screenshot, clip or recording described.
Good fit when
- A text model must read the labels on an image.
- A clip needs analysis of both visuals and the soundtrack.
- The structure, rhythm and instrumentation of a music piece need describing.
- Different providers are connected for vision and hearing.
Not a fit when
- Files must not leave the machine, because media goes to the chosen provider.
- Continuous monitoring or playback is needed, which the server does not do.
- FFmpeg is unavailable but video or audio processing is required, since both depend on it.
Example request
Look at this screenshot and read the labels on the buttons, then describe the audio in the attached clip.MCP
- Transport
- stdio
- Authentication
- API key
Limitations
Distributed as source code: Git, Python 3.11 or newer and uv are required, and there is no packaged installer. Video and audio need a system FFmpeg. The self-test sends tiny test media to the provider and may cost money. The last commit dates from August 2026. The provider applies its own retention and billing rules to your files.
How to disable. Remove the server entry from the client settings and uninstall the `cove-sensory-mcp` package. Configuration and credentials remain until you explicitly request removal.
FAQ
How do I connect Cove Sensory MCP to Claude Code?
Run `uv run cove-sensory-mcp print-config --client claude-code`, copy the printed stdio entry into the client settings and restart it. Then ask the agent to call `sensory_setup_guide`.
Where is the API key stored?
In the operating system credential store. An `env:VARIABLE_NAME` reference stores only the variable name, which must be set in the process that launches the MCP.
Do I need all five capabilities?
No. MiniMax-M3 covers images and video visuals without hearing, so connect another provider for sound. In the MCP self-test, request only configured capabilities.
Which file paths are allowed?
Absolute paths inside directories added with `cove-sensory-mcp configure paths`. URLs must be direct HTTPS media links; localhost and redirects are blocked.
What if setup is interrupted?
Run `uv run cove-sensory-mcp configure` again and select the same provider. The wizard detects the saved entry before asking for keys and offers to continue.
Discussion
No questions or reviews yet.
Related
MCP serversClaude Video Vision
claude-video-vision
An MCP server and Claude Code plugin that let the agent watch video: frames via ffmpeg and audio transcription via Gemini, local Whisper or OpenAI
A local MCP server for editing and analyzing images: the agent crops frames, finds objects and reads text from photos
The official remote MCP server of Kling AI: an agent makes videos and images from text and from a picture right in the work chat
MCP server for image generation and editing: Gemini, OpenAI and Seedream with automatic prompt enhancement
An MCP server for Civitai: model and LoRA search, images with prompts, download commands, trend analysis, bilingual output
WaveSpeed MCP server: an agent runs image, video, audio and 3D models, reads each model's input schema and quotes the price first
README in short
The server is meant as a sensory layer limited to answering an agent's question about an explicitly supplied file. The setup wizard checks key access before asking for credentials and can resume an interrupted setup.