skycam

The camera records the sky continuously. Video goes to openskycam (SkyVision: a Pi uploads 60-second segments to Cloudflare R2), and the Mac mini runs detection workflows over it. You reach all of it through the openskycam MCP server: guide, footage, workflows, runs, images, buckets, models. Each takes an action.

The tools load on first use (search for openskycam if they aren't listed). The server's rules come with its instructions, and they bind: read guide before acting, times are UTC, preview before any delete, verify every change by reading it back.

Triggers

• "what's the sky look like", "show me the camera", "anything overhead": a photo of now
• "what did the sky look like at sunset / at 3 this morning": a photo at that time
• "did the camera catch anything", "any planes / birds / meteors today": detections from runs
• "send me the video from 9pm": footage download
• "run the plane detector on last night": a workflow run
• "is the camera working": footage, action cameras

A photo of right now

1. footage with action cameras. Take newest_recording for camera sky01. It is one to three minutes behind real time; say so if it matters. If receiving is false, the camera or its uploader is down. Say that rather than sending an old frame.
2. footage with action photo, camera: "sky01", and at about 30 seconds after newest_recording, since the time must fall inside a recording. It takes about 5 seconds. If only a photo_id comes back, call photo again with that photo_id.
3. Look at the preview it returns and describe what is actually there: cloud cover, light, anything in frame. Report captured_at in Eastern time, not UTC.
4. To show it, download the full frame into this conversation's folder and send the file:

   curl -fsSL -o "sky-$(date +%Y%m%d-%H%M%S).jpg" "<full_size.url>"

   Then send_attachment(file_path=<that file>, caption=<one line>). The file is 3840x2160, about 1 MB.

Never paste a download link into a chat. Links grant access to their file for an hour, and the server's rule is to give them only to the user. Download and send the file instead.

A past moment

Use footage with action days for the month, then action list with from/to around the moment. Then take a photo inside a listed recording, between start and start + seconds. For a series of frames, such as a sunset or a before and after, take several photos. For many frames at a fixed interval, build a workflow instead.

What the models found

• images, action search: filter by class, min_confidence, camera, from/to, bucket_id or run_id. Newest first.
• images, action view: shows up to four images at 768 px, with their detections drawn in.
• images, action link: full size. Download it and send the file, as above.

Say what was detected and when, and how confident the model was. Don't claim a detection is a specific object when the class is generic.

Workflows and runs

Follow guide with build-a-workflow and verify-changes. The order is fixed: runs action test before start, then watch until the run finishes. Each watch waits at most 45 seconds, so call it again. The Mac mini runs one workflow at a time: a new run waits behind an unfinished one, and waiting_behind says which. Starting a long run is the operator's call; check before starting one someone else asked for.

Changes and deletes

Create, update or delete workflows, buckets, models or devices only when you were asked to, in so many words. Edits carry the revision you last read; on a conflict, read again and reapply. A delete returns a preview first. Confirm only what was asked for, then read it back and report what you observed.

When it's broken

• receiving: false, or no recording in the last few minutes: the camera or its Pi uploader is down.
• Photos never finish: runs, action runtimes. The Mac mini processor (launchd org.openskycam.workshop) must have checked in recently.

Tell the person what is down. Don't describe a frame you didn't get.