Home Assistant integration

Glad it worked for you! My plan is kind of a hybrid approach-using the HACS integration for toggling the schedule presets or individual camera's actions, and then adding an input boolean from Home Assistant for each camera in the Actions section of Home Helper. Then I can use those input booleans to trigger automations in Home Assistant. This is a big win for me since I'm away from home 5-6 days at a stretch for work, so the only way I have to edit stuff in Home Helper is via screen sharing, which tends to be kind of slow, while I can edit Home Assistant automations from my phone.

Claude and I updated this today to fix some stability stuff I'd experienced, namely that the integration silently failed if I restarted my Mac or Security Spy. I also changed the way cameras are added, since I'd noticed that they'd be added multiple times if I swapped cameras around in Security Spy. I also got the "reconfigure" button working in Home Assistant, so if you change your Security Spy password you don't have to remove the whole thing from Home Assistant and add it again. Anybody that runs across this, feel free to reach out with suggestions! The Security Spy API exposes other stuff that could be useful, like schedule overrides, it's just not stuff I use with Home Assistant, so I didn't add it. I'd be glad to put it in there if it could be helpful to someone.

You got it! The "Enable" toggle switch will be available when you get the integration going now, but it will be hidden by default since it's using a POST request instead of a GET request like the rest of the integration-and also because I'm afraid I'd fat-finger the toggle in the HA app. Just click where you see the "+1 disabled entity," hit the gear wheel, and then toggle the "enabled" switch. Home Assistant should then tell you "The enabled entities will be added to Home Assistant in 30 seconds." (This is confusing to type, because the entity is called "enabled," but you also need to enable the enabler. Enabling you to enable the enabler really makes me sounds like a bad person.) I think you'll see what I mean when you get there.

I'm such a GitHub rookie, I didn't realize I had issues and discussions disabled, that should be fixed now.

Your Roku setup looks like it'd be slick! I don't have a (running) Roku, I just got a new Apple TV because my old one was being flaky about HomeKit (which was the reason I wanted to fix this Home Assistant integration in the first place) but I kinda want to break out my old Roku just to try your viewer.

It's not that I don't think you have a great use-case for that switch, I just mean that it's doing more than just toggling a schedule preset. To do this in the UI, you have to change a setting, then save, and the camera disappears from the list, like it's not even counting towards your license anymore. The mute switch would solve that part of it, or maybe a whole screen privacy mask? And not to make you paranoid, but you've still got a camera inside that's got power and is on your network unless you're also cutting power to it.

I'd be interested in seeing your app. I built myself a little flask app running on the Mac that waits for the "Actions" tab in Security Spy to fire a shell command on "human moves," watches for new still images to drop, then sends some of those images along with a bunch of context and reference images up an Anthropic API call with a prompt like "you're reviewing my cctv footage because I'm driving, let me know if I should pull over to review it..." The Flask app then sends me an iMessage so Siri can tell me what's up without me having to watch the footage myself. It's clunky, but it works, and I'm not itching to pull over to watch a video that just ends up being somebody cleaning up after their dog in my yard.

I inadvertently made a breaking change on this in version 1.3 that could orphan entities for some folks trying to fix bug in the integration where moving or disabling cameras in Security Spy could create duplicate cameras in Home Assistant. The fix is out now with 1.3.2, and should work without orphaning entities for folks that added the new integration at any point. Holler at me here or on GitHub with any issues.

@photonclock , I tried the larger version of that model, and it was a learning curve I didn't know a was about to navigate! It didn't work out for what I wanted to use it for, but it led to me figure out a better way of using qwen3.5:9B so that it's usable on the Mac at the same time as Security Spy. My last message from it while I was trying to stress-test it with what ended up being 21k tokens was

"Josh in backyard and front porch at 10:39-10:41 AM wearing plaid shirt with blue denim overalls, walking from backyard through east side yard to front porch, standing on porch and in yard for approximately 45 seconds across Back East, East corner S, Front Door, Front Floodlight, Porch East, and Porch West cameras. No concern - known person (homeowner) moving around property normally."

I did get this gem out of FasVLM when I was running it though-not totally accurate, but I think the 'Possum would be proud.

The biggest problem with FastVLM (the 7B version I used) was the long text prompts I fed it to try to get it to describe the action in a scene. It would end up repeating some version of the prompt almost verbatim. It was great at "Describe this image" but not so good at "You are a security officer reviewing CCTV footage...." It evidently was trained from Qwen2, and the Qwen models have come a long way since then. It turns out a picture is better than a thousand words for baby LLMs too!

I really do need to make a good write up on this at some point, but it's a real moving target. In broad strokes, when Security Spy's actions trigger, which for me is just set for triggering on "Human Moves," the shell command it fires starts a folder watcher for new still images to come in. When they stop coming in, or two minutes after the first one comes in, the flask app grabs eight of them and sends them and a huge prompt that includes reference night/day images and pictures of known people, which is how it named me. The big downside to that is just suggesting it compare an image of a known person to a new, unknown person makes it into a matching game for it. The big brain LLMs aren't immune to this either, they try really hard to match people, so I'm thinking about removing that whole section. Anyway, it takes all of that, is told explicitly to use certain keywords, and send its response back in a format that's easy for Siri to read me through CarPlay. It takes an API call about 30-40 seconds and the 9b Qwen about 160-170.

I'm working on another version now that will run the images through in series with decision trees where the bot just fills out a json template. "Are there people in this image? Yes/no" kind of logic, followed by getting it to describe them, whether they're wearing a uniform or not, etc, and then to a third stage that determines where they entered the scene, where they left it, and how long they were there. The last stage will ask it to compare the image immediately before the person entered and immediately after it left and make note of anything taken, left, or altered. Then all of that info gets pasted in a template and sent off as an iMessage. The hope is that it'll be faster and lighter-weight than a giant prompt with a lot of decision making by an LLM.