LLM Vision is a Home Assistant integration that uses multimodal large language models to analyze images, video files, live camera feeds and Frigate events, built for smart-home owners who want their cameras to describe what they see rather than simply report that motion occurred.

What it is

LLM Vision is an open-source, Apache-2.0 licensed Python integration for Home Assistant, distributed as a HACS integration and available in the default HACS repository. It sits in the Home Assistant ecosystem and sends camera imagery, video and event data to multimodal large language models, returning natural-language answers, descriptions and structured data that Home Assistant can act on. It supports hosted providers such as OpenRouter, OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure and Groq, as well as locally hosted options including Ollama, Open WebUI and LocalAI, plus any service exposing OpenAI-compatible endpoints.

The problem it addresses is the gap between a camera alert and an actual answer. A standard camera notification tells a household that something moved; LLM Vision instead answers questions and produces descriptions of images, video files, live feeds and Frigate events based on a prompt the user writes. It remembers people, pets and objects, keeps a timeline of camera events for the dashboard, and updates sensors from data extracted out of camera streams. The specific thing it replaces is the undifferentiated camera notification and the manual scroll back through footage that follows it.

Key capabilities

• Analyzes images, video files, live camera feeds and Frigate events, answering questions and generating descriptions from a user-supplied prompt.
• Connects to OpenRouter, OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure, Groq, Ollama, Open WebUI, LocalAI and any provider with OpenAI-compatible endpoints.
• Remembers people, pets and objects across analyses.
• Keeps a timeline of camera events, displayable with an optional Timeline Card, which can also be queried through Home Assistant Assist.
• Updates Home Assistant sensors with data extracted from camera streams, images or videos, so detections can drive automations.
• Ships with a blueprint that delivers camera event notifications intelligently summarized by AI.
• Stores snapshots in the more secure /media folder; Home Assistant Container users may need to mount a folder to /media in their container settings.

Who uses it and how

• Home Assistant households with one or more cameras that want notifications describing what was seen, using the bundled blueprint instead of plain motion alerts.
• Users running Frigate, whose events are one of the input types the integration analyzes.
• Self-hosters who prefer local inference, pointing the integration at Ollama, Open WebUI or LocalAI rather than a commercial cloud API.
• Dashboard builders who install the Timeline Card to review past camera events, or ask Assist about them conversationally.
• Automation authors who consume the sensors the integration updates with data extracted from streams, images and videos.

Getting started

Install LLM Vision from HACS, restart Home Assistant, add the integration under Settings / Devices & services, set up the /media folder, and then press "Add Entry" to configure an AI provider. Detailed provider setup lives in the documentation at llm-vision.gitbook.io, with the project site at llmvision.org.

How it compares

No list of paid products that this project replaces is provided, and the facts name no directly comparable camera-analysis integration. Its nearest named relationships are with the LLM providers it calls and with Frigate, whose events it consumes; on the evidence here, it stands alone in this registry.

When to use it — and when not

Adoption means running Home Assistant, installing through HACS, arranging a /media mount for snapshots where needed, and supplying an AI provider — either a cloud API key with its associated cost, or local hardware running Ollama, Open WebUI or LocalAI. Anyone outside the Home Assistant ecosystem, or unwilling to operate or pay for a model provider, should not choose it. The main limitation visible in the facts is that the repository README is largely a pointer: full setup instructions and provider documentation live off-repository at llmvision.org and the GitBook docs.