Skip to content

Record a sky camera around the clock, keep the footage on your own disk, and find what flew past.

3.7 KB
Raw Blame

title: Intro sidebar: Intro description: openskycam records a sky camera around the clock, keeps the footage on your own disk, and finds what flew past.

openskycam keeps every frame your sky camera encodes: a relay copies the stream into 60-second MP4 files without re-encoding. The files live on a disk you own. Cloudflare Workers catalog them, serve a website and an API, and hand frames to a processor that runs detection models. Packages and cloud resources keep the project's first name, SkyVision (skyvision-dashboard, skyvision_processor, the skyvision-catalog database).

A meteor in the review view: the detection model labeled it Meteorite, and a reviewer confirms or corrects the label

IP camera ──RTSP──▶ relay ──HTTPS PUT──▶ ingest Worker ──Tunnel──▶ footage store
 H.264 or H.265     Linux host,          checks, names,              your disk,
                    60 s MP4s            catalogs in D1              rolling 40 days

website + API Worker ◀── Cloudflare Access (people), API keys (scripts, agents)
        │
        └── jobs over HTTPS ──▶ processor: decodes frames, runs YOLO models, tracks motion
                                (macOS with Apple GPU, Linux with NVIDIA GPU, or CPU)

What you can do with it

  • Browse and download footage. Pick a day, scrub through hours, skip by 1 to 30 seconds, loop a few seconds, zoom into the 4K frame, and download the original file. Footage stays for 40 days.
  • Find things in it. A Workshop workflow samples frames from a time window, runs one to four detection or classification models, and routes each frame to a bucket by class, score, count, time of day, or motion. Workflows can run on a schedule, for example every night.
  • Tell planes from stars. Boxes from the same object are linked across a recording into a track. Each track is moving, stationary, or brief, so a workflow can keep moving planes and set fixed lights aside.
  • Build training data. Review detections, accept or reject them, correct their boxes, and freeze a dataset version. A version records its images, labels, tracks, classes, and splits, and a script can download it.
  • Hand it to an agent. Everything the website does is also an API call under /v1, and an MCP server gives AI agents the same actions with guides on how to use them. An agent can run the setup, operate it day to day and change its code (Run it with an agent).

What it is made of

Part Runs on Code
Relay A Linux host near the camera, such as a Raspberry Pi external/relay/
Ingest Worker Cloudflare package/worker/
Footage store A macOS or Linux host with a large disk external/store/
Website and API Worker Cloudflare package/dashboard/
Processor A macOS or Linux host, GPU optional processor/
Catalog schema Cloudflare D1 db/migrations/

The store and the processor can be the same machine. Nothing in Cloudflare holds footage: the Workers reach the store through a Cloudflare Tunnel.

Where to go next

  • How it works
    The path of a recording, and of a workflow run, including what happens when a part is down.
  • Requirements
    Cameras, hosts, disk sizing, and the Cloudflare account.
  • Deploy your own
    Every step from a clone to video arriving.
  • Run it with an agent
    What an AI agent can do over MCP, and how it runs the setup for you.
  • Configuration
    Every setting and its default.
  • Workflows
    Sample frames, run detection models, and sort what they find into buckets.
  • API
    Keys, scopes, and the /v1 API that does everything the website does.