What is Frigate?
Frigate records IP cameras and runs object detection on the video locally, so it can tell a person or a car from moving branches and record or alert only on what matters. A cheap motion detector decides where to look, and the object detector runs only on those regions. It records around the clock or by event, re-streams over RTSP so each camera serves one connection, and offers WebRTC and MSE live view.
It was built for Home Assistant, with its own integration, and publishes events over MQTT for anything else. Version 0.18 added full configuration from the web UI, profiles, and a GenAI chat that answers questions about camera activity through llama.cpp, Ollama or cloud providers such as Gemini and OpenAI; GenAI descriptions are off by default.
The code is MIT-licensed and written in Python, but the Frigate name and logo are trademarks of Frigate, Inc. and not covered by the licence. Frigate+ is an optional $50-a-year subscription for better detection models, including up to 12 fine-tuned on images you upload.
Is Frigate worth it?
Frigate turns ordinary IP cameras into a system that knows the difference between a person at the door and a tree in the wind, without sending footage to anyone's cloud. It has 37,569 GitHub stars, and 0.18.2, released on 15 October 2026, is current. Detection runs on your hardware. The default is a single CPU detector, which the docs say is not recommended for general use; the supported list covers Coral Edge TPU, Hailo-8 and 8L, Intel OpenVINO on 6th-generation or newer chips, NVIDIA through TensorRT, AMD ROCm and Apple Silicon through a host-side client, with Rockchip, MemryX and others community-supported. Detector types cannot be mixed for object detection.
Footage of your home is about as personal as data gets, and the defaults need attention. Frigate serves an authenticated UI and API on port 8971 and an unauthenticated one on port 5000, where, the docs say, every request gets access equivalent to the admin role. Port 5000 belongs inside the Docker network or a trusted LAN, never on a forwarded port. An admin password is generated on first start and printed in the logs, and login rate limiting is a setting you have to add. GitHub lists 38 security advisories, 32 of them published in 2026, including five critical remote code execution flaws; two were fixed in 0.18.2, one of which let a viewer-role user run commands as root. Running anything older than 0.18.2 is hard to defend.
What leaves the house is your decision. Object detection, face recognition and licence plate recognition run locally. GenAI descriptions and chat stay off until you enable them; a cloud provider then receives thumbnails or frames from your cameras, while llama.cpp or Ollama keep it local. Frigate+ ($50 a year, extra fine-tunes $5) works on images you choose to upload and annotate; its base models are trained on images from other Frigate+ users, uploading is optional, and models trained during a subscription stay downloadable after it lapses.
Upgrades take care. 0.18.0 carried a long list of breaking changes, from the GenAI config to the zone and mask format, and the notes advise copying the config and frigate.db first; Intel GPU stats now need Linux kernel 6.5 or newer, and the DeGirum detector was removed after the company ceased operations. Even with the new settings UI, getting streams, hardware decoding and go2rtc right takes an evening of reading.
Where it earns its place
A homeowner with four PoE cameras and Home Assistant on a mini PC
Get a phone alert when a person, not a cat or a shadow, enters the driveway
Frigate runs object detection on a USB Coral or the mini PC's Intel GPU through OpenVINO, a driveway zone limits alerts to that area, and the MQTT event triggers a Home Assistant notification with a snapshot. No per-camera cloud fee.
A small shop owner with six cameras covering the till and stockroom
Keep 24/7 recordings and jump straight to moments with people in frame
Continuous recording with longer retention for clips containing detected objects, a review page that lists alerts per camera, and viewer-role accounts limited to specific cameras for staff. Access from home goes through port 8971 behind a VPN.
A Home Assistant user with an NVIDIA GPU and a local Ollama server
Ask what happened at the front door this afternoon in plain language
Enable GenAI review descriptions and chat with Ollama as the provider, so frames are described locally. The chat agent can search tracked objects and summarise recent activity, and no images reach a cloud provider.
Before you commit
- Setup
- Docker on Linux with shared memory and device passthrough for the accelerator; pick the image variant that matches it (-tensorrt, -rocm, -rk and others). The admin password is printed in the logs on first start, and cameras can be added from the settings UI.
- Integrations
- Home Assistant integration, MQTT topics for events and profiles, an HTTP API, RTSP re-streaming through go2rtc, WebRTC and MSE live view, and GenAI providers: llama.cpp, Ollama, OpenAI-compatible servers, Gemini, OpenAI and Azure OpenAI.
- Data and privacy
- Recordings, snapshots, faces and plate reads stay on your disks. Cloud GenAI providers receive thumbnails or frames only when you enable those features. Port 5000 has no authentication and grants admin-level access.
- Cost at scale
- No licence fee. Costs are disk for recordings, an accelerator (a Coral, Hailo or GPU) sized to the number of cameras, and optionally Frigate+ at $50 a year. A single per-camera hardware figure is not published; it depends on the detector.
- Limits
- Only one detector type can run object detection at a time; the CPU detector is not recommended; Intel GPU stats need kernel 6.5 or newer; FFmpeg 5 support is due to be removed in 0.19.
What’s good
- Object, face and licence plate recognition run on your own hardware
- Wide accelerator support, from a USB Coral to NVIDIA, Intel, AMD, Apple Silicon and Hailo
- Home Assistant integration and MQTT events for any other system
- Full configuration from the web UI since 0.18, with YAML still supported
What’s not
- Unauthenticated admin-level API on port 5000 that must never be exposed
- Five critical remote code execution advisories, two fixed only in 0.18.2
- CPU-only detection is not recommended; plan for a Coral, Hailo or GPU
- Major releases bring breaking config changes and need a backup first
Best for
Home Assistant users and small sites with several IP cameras who want people and vehicle alerts without a cloud camera subscription, and can run a small server with an accelerator.
Frigate: common questions
- Is Frigate worth it?
- Camera recording that keeps footage and detection inside your house and spots people instead of shadows, provided you give it a detector chip or GPU and patch it promptly. We rate it 4.4 out of 5.
- How much does Frigate cost?
- Free and open source under MIT (name and logo trademarked). Optional Frigate+ at $50/year for improved models and up to 12 fine-tunes; extra fine-tunes $5 each.
- What are the downsides of Frigate?
- Unauthenticated admin-level API on port 5000 that must never be exposed; five critical remote code execution advisories, two fixed only in 0.18.2; cPU-only detection is not recommended; plan for a Coral, Hailo or GPU; major releases bring breaking config changes and need a backup first.
- Who is Frigate best for?
- Home Assistant users and small sites with several IP cameras who want people and vehicle alerts without a cloud camera subscription, and can run a small server with an accelerator.
- How much does Frigate cost for a team or in production?
- No licence fee. Costs are disk for recordings, an accelerator (a Coral, Hailo or GPU) sized to the number of cameras, and optionally Frigate+ at $50 a year. A single per-camera hardware figure is not published; it depends on the detector.
- Is Frigate safe for company data?
- Recordings, snapshots, faces and plate reads stay on your disks. Cloud GenAI providers receive thumbnails or frames only when you enable those features. Port 5000 has no authentication and grants admin-level access.
Answers checked Nov 18, 2026.