Turn any low-power ARM device into a professional NVR in 60 seconds. Single binary, zero dependencies, no cloud required. Runs on low-power ARM devices with a 512MB memory budget.
Self-measured — methodology and full curves in Performance.
Surveillance grid: multi-camera live view (WebCodecs / MJPEG) with health scores and AI overlay
Recordings: per-camera timeline with AI event markers; list and timelapse views available
Cameras: discovery scanning, drag-and-drop groups, start/stop, and push-out relay management in one place
curl -fsSL https://raw.githubusercontent.com/Mi-Bee-Studio/MiBeeNvr/main/install.sh | sudo bashDownloads the binary, creates a system user (nvr), generates a config, installs a systemd service, and starts it. Data directory: /var/lib/mibee-nvr. Open http://localhost:9090 for the Web UI.
docker compose --project-directory . -f deploy/docker/docker-compose.yml up -dTo store recordings on an external drive, edit the volume mount and set NVR_DATA_DIR=/data to match (see the commented block in docker-compose.yml).
Other options: pre-built binary, build from source
# Pre-built binary (AMD64 / ARM64 / ARMv7)
wget https://github.com/Mi-Bee-Studio/MiBeeNvr/releases/latest/download/mibee-nvr-amd64
chmod +x mibee-nvr-amd64
./mibee-nvr-amd64 init --password yourpassword
./mibee-nvr-amd64 -config mibee-nvr.yaml# Build from source
git clone https://github.com/Mi-Bee-Studio/MiBeeNvr.git
cd MiBeeNvr && make buildFor detailed setup, see Getting Started.
- Single Binary: Zero dependencies, embedded Svelte 5 SPA,
CGO_ENABLED=0 - Low-Power ARM Ready: Runs on 1GB RAM with a 512MB memory budget (validated on Raspberry Pi 3B as the minimum baseline)
- Plug-and-Play Auto-Discover: ONVIF cameras joining the LAN are enrolled automatically in the background — no manual scan; authenticated ones wait as pending activation until you supply credentials
- IP Self-Healing: Cameras that change IP (Wi-Fi roaming across APs) are auto-relocated by serial number and reconnected
- Adaptive Recording: Compressed-domain motion analysis drops calm scenes to a timelapse-grade baseline and resumes full rate on motion or loud audio — 75–98% less disk on static scenes, zero frames lost at the switch (details)
- AI Detection, Two Layers: Browser-side ONNX Runtime Web with WebGPU acceleration — no inference server needed; optional server-side MiBeeVision companion analyzes recordings, writes events back and drives semantic triggers
- Modern Web UI: Dark/light themes, Chinese / English / Russian, responsive; WebCodecs H.265 playback
- No Cloud Required: Self-hosted, no subscriptions, no vendor lock-in — remote access via Tailscale / Cloudflare Tunnel / WireGuard today, with a built-in encrypted remote mode on the way
In most surveillance scenes nothing happens 95% of the time — yet continuous recording writes every frame at full bitrate. Adaptive recording analyzes the compressed stream on the NVR side (P-frame sizes vs a rolling baseline; no decoding involved) and drops calm scenes to a timelapse-grade baseline, resuming full rate the moment activity appears. A GOP pre-buffer is flushed at the switch, so the sparse→full transition loses zero reference frames — no corruption, no playback gaps.
- 75–98% less disk on static scenes — field data: a 2K H.265 camera went from ~2700 MB/h to ~688 MB/h including active spans
- Five wake-up signals: bitrate spike · loud audio (G.711 dBFS) · pixel-domain CV sampled at ~1fps from the sub-stream (for smart-codec cameras whose bitrate never moves) · camera-side ONVIF motion · a semantic trigger API driven by MiBeeVision or any AI / home automation
- Tiered recording: the sub-stream records 24/7 as a never-miss baseline while the main stream records events only — a combination still an open feature request elsewhere
- Offline tuning:
eval-replayreplays existing recordings through the gate — preview parameter changes without touching production
Full pipeline, tuning and telemetry: Adaptive Recording.
Self-measured across three hardware tiers — methodology, judging criteria, and full resource curves are published in the benchmark report. Default configuration throughout; no manually-enabled optimization switches.
| Hardware | Load | Result |
|---|---|---|
| Any (idle) | — | 38 MB RSS, 122 ms cold start |
| x86 desktop (i5-12600KF) | 68 cams · 226 Mbps recording, +32 concurrent viewers | 24% CPU, API p50 < 11 ms, zero failures |
| RK3399 NAS (low tier) | 28 cams recording | 20–23% CPU (storage-bound, eMMC) |
| Banana Pi M5 (production) | 18 cams, 80% iowait storm | health API p50 13.5 ms |
| Protocol | Direction | Notes |
|---|---|---|
| RTSP | Camera → NVR | H.264/H.265/MJPEG |
| HTTP JPEG | Camera → NVR | Snapshot streaming |
| ONVIF | Camera ↔ NVR | Discovery, PTZ, stream URI, MotionAlarm events |
| Xiaomi (CS2 + TUTK) | Camera → NVR | CS2 P2P + legacy TUTK (7 models), cloud auth |
| GB/T 28181 | Camera → NVR | SIP platform, auto-enrollment, audio + talk, PTZ, device-side playback (production support since v0.14) |
| SIP doorbell | Door station → NVR | Doorbird/Akuvox/Hikvision intercoms: ring popup, two-way G.711 talk |
| SRT / RTMP (ingest) | Publisher → NVR | Push-in: remote publishers, low-latency transport |
| WHIP | Publisher → NVR | WebRTC push from browsers, phones, OBS 30+ (H.264 + Opus) |
| HLS | NVR → Viewer | On-demand with LL-HLS support |
| WebRTC (WHEP) | NVR → Viewer | Sub-second latency |
| HTTP-FLV | NVR → Viewer | Browser-friendly streaming |
| DLNA / UPnP AV | NVR → TV | LAN media server — TVs browse and play the recording library |
| RTSP (output server) | NVR → External | Every camera re-served at rtsp://<nvr>:8554/<camera-id> |
| RTMP / RTSP (relay) | NVR → External | Native Go push-out to remote targets |
Full index with one-line descriptions: docs/en/README.md. Most-visited entry points:
- Getting Started · Configuration · CLI · API Reference
- Deployment (incl. NAS guides: unRAID / fnOS / Synology / QNAP / iStoreOS / ZSpace; native deb/rpm packages) · Remote Access
- Troubleshooting · Upgrade Guide
Version numbers, configuration syntax, and CLI flags are authoritative only in this repository's releases and docs/en/. Content published elsewhere — reprinted tutorials, "reviews", SEO aggregators — is not reviewed by us and frequently contains invented version numbers and benchmarks. If something you read elsewhere does not match the docs, trust the docs and please open an issue.
WeChat official account: 蓝宝石的傻话 (release announcements, in Chinese)
- Run
make lintbefore submitting (golangci-lint v2;make lint-installfirst) and add tests for new features - Web SPA: Node 22,
cd web && npm run build && npm run test - Write clear commit messages
Project layout — cmd/mibee-nvr/ CLI entry, internal/ core packages, web/ Svelte 5 SPA, deploy/ service files, docs/ bilingual documentation, tests/ + e2e-tests/ integration and Playwright tests. Details in the architecture overview.
Thanks to everyone who has contributed:
AGPL-3.0-only © Mi&Bee Studio, with a linking exception for pkg/:
- Using MiBee NVR (running it, recording cameras, watching streams): no obligations — use it freely, including commercially.
- Modifying and distributing it: your modified version must be released under AGPL-3.0.
- Building on
pkg/in your own program: allowed under the linking exception; your program's license is your choice. - Separate programs talking to a running NVR over its HTTP/WebSocket APIs: never affected by the license.
Releases up to v0.10.1 were MIT-licensed; that continues to apply to those versions. See NOTICE and CONTRIBUTING.md.