diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..9a3d21a
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,142 @@
+# AGENTS.md
+
+Notes for coding agents working in this repo, and for agents writing code
+against the Sonilo Python packages.
+
+## What this repo is
+
+Three PyPI packages in one repo:
+
+| Package | Path | What it is |
+|---|---|---|
+| `sonilo` | `.` (`src/sonilo`) | Python client for the Sonilo API. Sync (`Sonilo`) and async (`AsyncSonilo`) clients, Python 3.9+, only depends on `httpx` |
+| `sonilo-cli` | `sonilo-cli/` | `sonilo` command line tool, built on `sonilo` |
+| `sonilo-video-kit` | `sonilo-video-kit/` | Helpers that generate a soundtrack, duck it under speech and mix it into a video with local ffmpeg |
+
+The TypeScript counterparts live in
+[sonilo-js](https://github.com/sonilo-ai/sonilo-js).
+
+## Layout
+
+```
+src/sonilo/_client.py            Sonilo: auth, base URL, timeout, transport
+src/sonilo/_async_client.py      AsyncSonilo, same surface with async methods
+src/sonilo/resources/*.py        one module per endpoint (text_to_music, video_to_sfx, dubbing, ...)
+src/sonilo/resources/tasks.py    tasks.get / tasks.wait and the parse_*_result parsers
+src/sonilo/errors.py             error classes, mapped from HTTP status and the API `code`
+src/sonilo/types.py              result dataclasses and TypedDicts
+src/sonilo/_version.py           __version__, sent in X-Sonilo-Client-Version
+sonilo-cli/src/sonilo_cli/       __main__.py (every command), login.py, credentials.py
+sonilo-video-kit/src/sonilo_video_kit/  generate, duck, mix, loudness, ffmpeg wrappers
+context7.json                    rules Context7 serves to coding assistants (checked by sonilo-cli/tests/test_context7.py)
+examples/                        runnable scripts
+```
+
+## Commands
+
+CI (`.github/workflows/ci.yml`) runs on Python 3.9 and 3.12 with ffmpeg
+installed:
+
+```bash
+pip install -e ".[dev]" -e "./sonilo-video-kit[dev]" -e "./sonilo-cli[dev]"
+pytest                    # sonilo
+pytest sonilo-video-kit   # needs ffmpeg and ffprobe on PATH
+pytest sonilo-cli
+```
+
+Tests mock HTTP with `respx`; none of them call the real API.
+
+## Making changes
+
+- Keep `sonilo` on Python 3.9 syntax (`Optional[...]`, `from __future__ import annotations`)
+  and on `httpx` as its only dependency.
+- Every resource has a sync class and an async class in the same module. Change
+  both, and cover both in tests.
+- New endpoint: add a module under `src/sonilo/resources/`, attach it in both
+  `_client.py` and `_async_client.py`, export new result types from
+  `src/sonilo/__init__.py`, add tests, document it in `README.md`, then add the
+  CLI command in `sonilo-cli/src/sonilo_cli/__main__.py` if a user would run it
+  from a terminal.
+- Use the API's field names (`video_url`, `variants_num`, `output_format`) as
+  the Python keyword arguments.
+- Tri-state options (`ducking`, `preserve_speech`) are sent only when not
+  `None`, so the server default applies when they are left unset.
+- Branch on the API's error `code`, never on message text.
+- Wrappers pass `client_name` / `client_version` (`cli-python`,
+  `videokit-python`) so their traffic is reported in the `X-Sonilo-Client`
+  header; direct SDK use leaves them unset.
+- `context7.json` rules are limited to 255 characters each and 50 in total;
+  `sonilo-cli/tests/test_context7.py` enforces it.
+
+## Releases
+
+Each package publishes to PyPI through OIDC Trusted Publishing when its tag is
+pushed. The workflow fails if the tag does not match the version in the
+package's `pyproject.toml`.
+
+| Package | Bump | Tag | Workflow |
+|---|---|---|---|
+| `sonilo` | `pyproject.toml` and `src/sonilo/_version.py` | `vX.Y.Z` | `publish.yml` |
+| `sonilo-cli` | `sonilo-cli/pyproject.toml` and `sonilo_cli/__init__.py` | `sonilo-cli-vX.Y.Z` | `publish-cli.yml` |
+| `sonilo-video-kit` | `sonilo-video-kit/pyproject.toml` and `sonilo_video_kit/_version.py` | `sonilo-video-kit-vX.Y.Z` | `publish-video-kit.yml` |
+
+`sonilo-cli` pins `sonilo>=X.Y,<X.(Y+1)`; raise that pin when the CLI needs a
+new SDK release.
+
+## Using the SDK
+
+Get an `sk-` key at https://platform.sonilo.com/dashboard/api-keys and set
+`SONILO_API_KEY`, or pass `api_key=`. The CLI also accepts the credential that
+`sonilo login` stores in `~/.config/sonilo/credentials.json`.
+
+```python
+from sonilo import Sonilo
+
+client = Sonilo()  # reads SONILO_API_KEY; base_url defaults to https://api.sonilo.com
+
+# Music streams back as m4a bytes; for wav or mp3 use generate_async(output_format=...)
+track = client.video_to_music.generate(video="clip.mp4", prompt="warm, hopeful")
+track.save("music.m4a")
+
+# Task-based endpoints: generate() submits and waits; submit() returns a task_id
+sfx = client.text_to_sfx.generate(prompt="glass shattering on stone", duration=5)
+sfx.save("sfx.m4a")
+
+from sonilo.resources.tasks import parse_sound_result
+task = client.video_to_sound.submit(video_url="https://example.com/clip.mp4")
+result = client.tasks.wait(task.task_id, parser=parse_sound_result)
+```
+
+`tasks.wait()` and `tasks.get()` default to the SFX parser; pass the matching
+`parse_*_result` for music, sound, dubbing, proofread and analysis tasks.
+
+Resources on the client: `text_to_music`, `video_to_music`, `text_to_sfx`,
+`video_to_sfx`, `video_to_video_music`, `video_to_video_sfx`, `video_to_sound`,
+`video_to_video_sound`, `video_analysis` (`analyze()`, returns a brief, not
+media), `audio_ducking`, `dubbing`, `proofread`, `tasks`, `account`.
+
+CLI commands: `sonilo text-to-music`, `video-to-music`, `text-to-sfx`,
+`video-to-sfx`, `video-to-sound`, `video-to-video-music`, `video-to-video-sfx`,
+`video-to-video-sound`, `audio-ducking`, `video-analysis`, `dubbing`,
+`proofread`, `tasks get|wait`, `account`, `usage`, `login`, `whoami`, `logout`.
+
+Video length caps (422 before any charge): music 360 s; SFX, sound and video
+analysis 480 s; audio ducking 360 s per input; dubbing and proofread 300 s.
+
+## Errors
+
+| Class | Status / code | What to do |
+|---|---|---|
+| `AuthenticationError` | 401 | Key invalid, revoked or expired; replace it or `sonilo login` again |
+| `TrialExhaustedError` | 402 `trial_exhausted` | Free runs for that service are spent; add a payment method at https://platform.sonilo.com/dashboard/billing. Retrying cannot help |
+| `PaymentRequiredError` | 402 `insufficient_balance`, `payment_required` | Top up, then retry |
+| `BadRequestError` | 400 / 413 / 422 | Bad parameters, file too large, over a length cap; `.detail` and `.errors` say which |
+| `RateLimitError` | 429 `rate_limit_exceeded` | Requests per minute: wait a full minute. Concurrency: wait for one of your running tasks to finish |
+| `GenerationError` | | A streamed `generate()` sent an `error` event mid-stream |
+| `TaskFailedError` | task `failed` | `.code`, `.task_id`, `.refunded`; failed generations are refunded |
+| `TaskTimeoutError` | | Still running and already charged; call `tasks.wait(err.task_id)` again instead of resubmitting |
+
+All of them extend `SoniloError`. `TrialExhaustedError` extends
+`PaymentRequiredError`, so catch it first. `client.account.services()` reports
+the remaining free runs (`trial`) and the account's `rpm_limit` and
+`concurrency_limit`.
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..43c994c
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1 @@
+@AGENTS.md
diff --git a/README.md b/README.md
index 558f584..e9d8ad0 100644
--- a/README.md
+++ b/README.md
@@ -26,12 +26,12 @@ then give it to the client either as an environment variable (recommended) or
 inline:
 
 ```bash
-export SONILO_API_KEY=sk_...
+export SONILO_API_KEY=sk-...
 ```
 
 ```python
 client = Sonilo()                  # reads SONILO_API_KEY
-client = Sonilo(api_key="sk_...")  # or pass it directly
+client = Sonilo(api_key="sk-...")  # or pass it directly
 ```
 
 Keep your key secret — use it only server-side, never commit it, and prefer the
@@ -48,7 +48,7 @@ track = client.text_to_music.generate(
     prompt="cinematic orchestral score",
     duration=60,
 )
-track.save("output.mp3")
+track.save("output.m4a")
 print(track.title)
 ```
 
diff --git a/context7.json b/context7.json
index 7686de1..dd165bf 100644
--- a/context7.json
+++ b/context7.json
@@ -27,7 +27,7 @@
     "Result media (.url) is a short-lived presigned URL, not the API's own domain \u2014 download it with the result's .save() helper; do not send the Authorization header to it.",
     "Catch AuthenticationError (401), PaymentRequiredError (402), RateLimitError (429) and TaskFailedError (status \"failed\") separately rather than one generic except \u2014 callers usually handle these differently. All extend SoniloError.",
     "video / video_url accept exactly one of the two, never both and never neither \u2014 validate before constructing a request.",
-    "Self-serve accounts start with free runs per endpoint (2 each for text-to-music, text-to-sfx, audio-ducking, video-analysis, proofread; 1 each for other video endpoints; none for dubbing). A first call succeeding is not proof billing works.",
+    "Self-serve accounts start with free runs per endpoint (2 each for text-to-music, text-to-sfx, audio-ducking, video-analysis, proofread; 1 each for other video endpoints; dubbing: one free 15 s preview). A first call succeeding is not proof billing works.",
     "Before a paid call, read client.account.services().get(\"trial\", {}) and degrade gracefully when a service's remaining is 0: that call raises TrialExhaustedError (402 trial_exhausted), which no retry fixes \u2014 ask for a payment method. trial may be absent.",
     "client.audio_ducking ducks an EXISTING music bed under an EXISTING voice track; nothing is generated. One of voice/voice_url, one of music/music_url. Voice may be audio or video (video returns a .mp4); music must be audio. Result: output_url, no stems.",
     "client.dubbing dubs one video into many languages in one call; languages: en, zh_cn, ja, ko, pt, pt_br, es, es_419, de, fr, it, ru, th, ar, tr, vi, id, ta, ml, kn, gu, pa_in, sd_in, hi; default zh_cn,es,fr. Billed per language; 1 free 15 s preview.",
diff --git a/examples/generate.py b/examples/generate.py
index ba9712f..92baec0 100644
--- a/examples/generate.py
+++ b/examples/generate.py
@@ -1,4 +1,4 @@
-"""Usage: SONILO_API_KEY=sk_... python examples/generate.py "lofi beat" 30"""
+"""Usage: SONILO_API_KEY=sk-... python examples/generate.py "lofi beat" 30"""
 import sys
 
 from sonilo import Sonilo
@@ -9,6 +9,6 @@
 with Sonilo() as client:
     track = client.text_to_music.generate(prompt=prompt, duration=duration)
 
-out = track.save("output.mp3")
+out = track.save("output.m4a")
 title = f' — "{track.title}"' if track.title else ""
 print(f"Saved {out} ({len(track.audio)} bytes){title}")
diff --git a/examples/sfx.py b/examples/sfx.py
index ab808c7..7045f9a 100644
--- a/examples/sfx.py
+++ b/examples/sfx.py
@@ -1,6 +1,6 @@
 """Generate a sound effect from text and save it locally.
 
-Usage: SONILO_API_KEY=sk_... python examples/sfx.py
+Usage: SONILO_API_KEY=sk-... python examples/sfx.py
 """
 from sonilo import Sonilo
 
diff --git a/sonilo-video-kit/examples/score_video.py b/sonilo-video-kit/examples/score_video.py
index 9df406e..0f23960 100644
--- a/sonilo-video-kit/examples/score_video.py
+++ b/sonilo-video-kit/examples/score_video.py
@@ -1,4 +1,4 @@
-"""Usage: SONILO_API_KEY=sk_... python examples/score_video.py clip.mp4 "upbeat, energetic"
+"""Usage: SONILO_API_KEY=sk-... python examples/score_video.py clip.mp4 "upbeat, energetic"
 
 Generates a soundtrack for a video and mixes it in locally with ffmpeg.
 Requires ffmpeg + ffprobe on PATH.