Car audio has two parts: what the car does, and where you are standing when it does it. The same engine heard from the pavement, the driver's seat or a hundred metres down the road gives you three different assets, and choosing the perspective is most of the work. These prompts name it every time, because library clips rarely say which one you are getting.
Every clip shows the exact prompt that produced it. The render script reads its prompts from the same file as this page. Copy one, change a word and generate your own.
petrol car engine starting from cold, starter motor cranking then catching and settling into a rough idle, exterior close
car door closing solidly, dense thunk with a rubber seal compressing and a latch click, exterior, street
car braking hard on dry tarmac, tyres screeching and locking into a stop, outdoor street, wide
car horn blasted once in city traffic, dual-tone honk with a street reverb tail, outdoor
car passing the listener at speed on a wet road, doppler engine and tyre roar sweeping past, roadside
interior of a car driving at motorway speed, muffled engine drone with tyre roar and wind, seamless loop
sports car revving hard and shifting up through two gears, throaty exhaust bark, exterior close
car indicator relay ticking steadily, dry mechanical clicks, interior close mic, seamless loop
car crashing head-on into a barrier, huge metal impact with glass shattering and debris scattering, street
car engine failing to start, starter motor cranking repeatedly and dying, battery weakening, exterior close
These models respond to physical description. Name the material, the action and the space.
petrol car engine starting from cold, starter motor cranking then catching and settling into a rough idle, exterior close
Describing the three stages (crank, catch, settle) gets you a complete event instead of an engine that is already running.
interior of a car driving at motorway speed, muffled engine drone with tyre roar and wind, seamless loop
"Interior" and "muffled" together put the listener inside the cabin. Without them the model renders the exterior every time.
car passing the listener at speed on a wet road, doppler engine and tyre roar sweeping past, roadside
The road surface changes the tyre noise more than the speed does: wet tarmac hisses, dry tarmac roars, gravel crunches.
car door closing solidly, dense thunk with a rubber seal compressing and a latch click, exterior, street
The rubber seal makes a modern car door sound expensive. Leave it out and you get a tinny older one.
Interior, exterior close and exterior distant are three different assets, and EQ does not convincingly turn one into another. The cabin is a small sealed space that removes the top end and adds its own resonance; distance removes the transients more than the treble. Write the perspective into every vehicle prompt. A missing perspective is the most common reason a car sound will not sit in a scene.
Most driving audio is two independent layers. The engine tells you what the driver is doing (idling, accelerating, labouring, shifting) and the tyres tell you what the road is. You can specify them separately: "revving hard" plus "on a wet road" gives you an aggressive car in the rain, a combination no single library clip is likely to have.
Indicator relays, seatbelt reels, handbrake ratchets, key fobs and latch clicks are small and instantly recognisable, and each costs one short render. A scene built from these plus one interior bed sounds more real than a single long drive-by, and gives you events to cut on. The indicator loop here is meant to sit under dialogue and mark time.
The models run on your machine, which changes the cost, the privacy and what you can automate.
Weights are fetched once. After that, generation, the library and the API all work with the network off. Everything lives in one SQLite file and one folder of WAVs under ~/.soundbox, and nothing is uploaded.
Soundbox runs an MCP server in the same process, with tools for generating and listing sounds. Generated files come back as absolute paths, so an agent can copy them into a project.
Every model's terms are shown before you download it, and non-commercial weights need an explicit confirmation. Soundbox is MIT. The licence of the model that made a sound governs what you may do with it.
claude mcp add --transport http soundbox http://127.0.0.1:17494/mcpcurl -X POST http://127.0.0.1:17494/api/generate \
-H 'Content-Type: application/json' \
-d '{"model":"audiocpp-stable-audio-3-sfx","prompt":"8-bit coin pickup blip","duration_s":1}'