tesla-clip-tools 0.3.0

Reusable primitives for IP-camera clip tools (Tesla, Wyze, Reolink, UniFi, Ring, Nest).

tesla-clip-tools

Reusable primitives for building tools that read Tesla TeslaCam, SentryClips, and SEI data — plus the same Source/Event/sampler/vlm/notify pipeline applied to Wyze, Reolink, UniFi Protect, Ring, and Nest.

Extracted from sentrytriage v0.2 → v0.3 so that future projects (planned: an FSD-Disengagement Studio, a Wyze/Reolink/UniFi port of sentrytriage) can share the same primitives without forking.

Install

pip install tesla-clip-tools                      # base
pip install 'tesla-clip-tools[sei]'               # adds protobuf+pandas for sei_dataframe()

Or with uv:

uv add tesla-clip-tools
uv add 'tesla-clip-tools[sei]'

Until the first PyPI release lands, install directly from GitHub:

uv add git+https://github.com/Raymondriter/tesla-clip-tools

What's inside

tesla_clip_tools/
├── sources/
│   ├── base.py          Abstract Source + Event dataclass
│   ├── tesla.py         SentryClips/<YYYY-MM-DD_HH-MM-SS>/{cam}.mp4 (HW3 4-cam + HW4 6-cam)
│   ├── wyze.py          Wyze SD-card / RTSP folder layout
│   ├── reolink.py       Reolink NVR folder layout
│   ├── unifi_protect.py UniFi Protect export layout
│   ├── ring.py          Ring downloads layout
│   └── nest.py          Nest / Google Home downloads layout
├── sampler.py           Frame extraction via imageio (bundled ffmpeg, no system dep)
├── vlm/
│   ├── base.py          Generic VLMBackend protocol parameterised on a Pydantic response_format
│   ├── openai.py        gpt-4o-mini default, structured outputs via beta.chat.completions.parse
│   └── ollama.py        Local Qwen2.5-VL etc. via the Ollama HTTP API + JSON-schema format
├── notify/
│   ├── base.py          Notifier protocol
│   ├── pushover.py
│   └── telegram.py
├── sei.py               SEI metadata extraction from Dashcam (Saved/Recent) clips
└── dashcam.proto        Tesla's published proto schema (vendored)

Generating the SEI protobuf module

For the SEI dataframe convenience function, you also need to generate the protobuf module once:

protoc --python_out=src/tesla_clip_tools src/tesla_clip_tools/dashcam.proto

iter_sei_payloads() works without protobuf — it just yields raw protobuf-encoded byte payloads.

Minimal example

from pathlib import Path
from tesla_clip_tools.sources.tesla import TeslaSource
from tesla_clip_tools.sampler import sample_keyframes
from tesla_clip_tools.vlm.openai import OpenAIBackend
from pydantic import BaseModel

class MyVerdict(BaseModel):
    is_interesting: bool
    why: str

source = TeslaSource(Path("~/Tesla/SentryClips"))
vlm = OpenAIBackend(model="gpt-4o-mini")

for event in source.discover():
    samples = sample_keyframes(event.cams, n_per_cam=4)
    captions = [f"[{s.cam} t={s.timestamp_seconds:.1f}s]" for s in samples]
    verdict = vlm.classify(
        images=[s.image for s in samples],
        captions=captions,
        system_prompt="Decide if this Sentry event deserves attention. Default to false.",
        response_format=MyVerdict,
    )
    print(event.id, verdict.is_interesting, verdict.why)

Design choices

• Generic VLM API. Every backend takes a response_format: type[BaseModel] so different consumers (sentrytriage, FSD-Disengagement Studio, future tools) define their own verdict schema without forking the backend.
• No state. No SQLite, no daemon, no notifications-orchestration. Those are application concerns; the library is just primitives. Use notify/ if you want one of the included notifiers, but it's optional.
• No system ffmpeg required for the read path. sampler.py uses imageio[ffmpeg] which bundles its own ffmpeg binary. (Tools that build filter graphs — the highlight reel in sentrytriage, e.g. — still need system ffmpeg.)
• Source-pluggable. tesla.py implements the canonical TeslaCam folder layout; wyze.py, reolink.py, unifi_protect.py, ring.py, and nest.py ship in 0.3. Add a new vendor by subclassing Source — the rest of the pipeline doesn't care.

Status

v0.3 — alpha, API may still change. Used in production by sentrytriage v0.3+ and fsd-disengagement-studio.

Changelog

See CHANGELOG.md for the version history. Releases are cut following the steps in RELEASE.md.

License

MIT.

CI

GitHub Actions runs ruff + pytest on Python 3.12 and 3.13 against every push and PR. See .github/workflows/ci.yml. The same tests also run as part of the workspace-level monorepo CI at C:\Dev\tesla\.github\workflows\ci.yml.