diff --git a/.claude/CLAUDE.md b/.claude/CLAUDE.md
index 71286b2..4bb31c3 100644
--- a/.claude/CLAUDE.md
+++ b/.claude/CLAUDE.md
@@ -34,6 +34,7 @@ Focused subpackages, each with a clear dependency footprint:
 | Module | Purpose | Heavy deps |
 |---|---|---|
 | `mixing.audio` | `Audio`/`AudioSamples`, fades, crop, concat, overlay, `loop_audio`, `duck_audio`, `find_audio_offset`, segmentation | `pydub`, `numpy`, `scipy` |
+| `mixing.audio.beats` | `beat_grid`/`BeatGrid` — beat/downbeat/onset analysis | `librosa` (extra `beats`, lazy-imported) |
 | `mixing.video` | `Video`/`VideoFrames`, crop/loop/speed, `replace_audio`, `normalize_audio`, `overlay_ambient_bed`, `ken_burns_*`, `concatenate_videos`, thumbnails, subtitles | `moviepy`, `opencv`, `pillow`, `looks` (vocabulary only — stdlib) |
 | `mixing.video.genai` | Google Vertex AI **Veo** generation | `google-genai` (extra `gen`) |
 | `mixing.transcript` | ElevenLabs **Scribe** STT (stdlib HTTP, cached), filler removal, SRT/prose | stdlib only |
@@ -124,7 +125,8 @@ Use the helpers, don't reinvent them:
 
 `pip install mixing[<extra>]`: `audio` (pydub/ffmpeg-python/soundfile),
 `widget` (Jupyter audio widget), `gen` (Veo/`google-genai`), `llm` (`aix`, for
-chapter titling + SRT translation), `clipboard` (`pyclip`), `dev`/`testing`.
+chapter titling + SRT translation), `clipboard` (`pyclip`), `beats` (`librosa`,
+for `mixing.audio.beats.beat_grid`), `dev`/`testing`.
 **ffmpeg must be on PATH** for most real work (`mixing.has_ffmpeg()` checks).
 
 ## Tests = guardrails
diff --git a/.claude/skills/mixing-audio/SKILL.md b/.claude/skills/mixing-audio/SKILL.md
index c5e320c..e0874b0 100644
--- a/.claude/skills/mixing-audio/SKILL.md
+++ b/.claude/skills/mixing-audio/SKILL.md
@@ -401,6 +401,27 @@ own `(AudioSegment, **kwargs) -> list[Segment]` callable as `strategy`. The
 `segment_by_silence/energy/self_similarity/speech_music` functions are exported
 too if you want to call one directly on a pydub `AudioSegment`.
 
+## Beat / downbeat / onset analysis (`mixing[beats]` extra)
+
+```python
+from mixing.audio import beat_grid, BeatGrid
+
+grid = beat_grid("song.mp3")  # -> BeatGrid; installs librosa lazily on first call
+grid.beat_times  # np.ndarray of beat instants, seconds, ascending
+grid.downbeat_times  # best-effort; empty array for the default (only) "librosa" backend
+grid.onset_env  # onset-strength envelope, one value per STFT hop
+grid.onset_hop_s  # seconds between onset_env frames (hop_length / sample_rate)
+grid.tempo_bpm  # estimated global tempo
+grid.to_dict()  # JSON-round-trippable summary (arrays -> lists; onset_env omitted)
+```
+
+`backend="librosa"` is the only shipped backend (ISC, commercial-clean); a
+stronger downbeat tracker can be added later behind the same `BeatGrid` shape.
+Needs `pip install mixing[beats]` (`librosa`) — `import mixing.audio` stays
+light without it; the import happens lazily inside `beat_grid` itself. First
+customer is muvid's footage-scoring layer (computes one grid on the master
+song, maps clips onto it via their known offset).
+
 ## Common recipes
 
 ```python
diff --git a/README.md b/README.md
index d42d334..a9b6cd4 100644
--- a/README.md
+++ b/README.md
@@ -128,13 +128,14 @@ pip install mixing[widget]     # interactive Jupyter audio widget
 pip install mixing[gen]        # Google Vertex AI Veo generation
 pip install mixing[llm]        # aix — chapter titling + SRT translation
 pip install mixing[clipboard]  # get file paths from the clipboard
+pip install mixing[beats]      # librosa — beat/downbeat/onset analysis
 ```
 
 ## What's inside
 
 | Subpackage | Highlights |
 |---|---|
-| `mixing.audio` | `Audio`, `fade_in/out`, `crop_audio`, `concatenate_audio`, `overlay_audio`, `loop_audio`, `duck_audio`, `find_audio_offset`, `find_segments`/`extract_segments` |
+| `mixing.audio` | `Audio`, `fade_in/out`, `crop_audio`, `concatenate_audio`, `overlay_audio`, `loop_audio`, `duck_audio`, `find_audio_offset`, `find_segments`/`extract_segments`, `beat_grid`/`BeatGrid` (extra `beats`) |
 | `mixing.video` | `Video`, `crop_video`, `loop_video`, `change_speed`, `replace_audio`, `normalize_audio`, `overlay_ambient_bed`, `ken_burns_video`/`ken_burns_film`, `concatenate_videos` (+ transitions: `crossfade_transition`, `fade_through_black`, …), `make_thumbnail`, `write_subtitles_in_video`, `SOCIAL_SIZES` |
 | `mixing.video.genai` | `generate_video` (Vertex AI Veo) |
 | `mixing.transcript` | `transcribe`, `remove_fillers`, `srt_for_media`, `words_to_srt`/`words_to_prose` |
diff --git a/pyproject.toml b/pyproject.toml
index 820243b..904e91c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -61,8 +61,9 @@ audio = [
     "ffmpeg-python",
     "soundfile",
 ]
-# Beat / downbeat / onset analysis (mixing.audio.beat_grid). librosa is ISC
-# (commercial-clean) and lazy-imported, so `import mixing.audio` stays light.
+# Beat / downbeat / onset analysis (mixing.audio.beats: beat_grid, BeatGrid).
+# librosa is ISC (commercial-clean) and lazy-imported, so `import mixing.audio`
+# stays light.
 beats = [
     "librosa",
 ]
@@ -146,7 +147,7 @@ install.macos = "brew install ffmpeg"
 install.windows = "choco install ffmpeg -y --no-progress"
 
 [tool.wads.ci]
-project_name = ""
+project_name = "mixing"
 
 [tool.wads.ci.commands]
 pre_test = []