Sound effects and foley from an AI agent

Published October 2026

Games, apps and videos need sounds that are not notes: a door slam on the cut, a whoosh under a title, footsteps on gravel, a UI blip, rain on a tin roof. doseedo_generate_sound makes them from a description, with Stable Audio Open on Doseedo's GPUs, as one MCP call or one POST /api/jobs. A one-shot hit, a loop cut to your tempo, or a texture, typically in 15 to 40 seconds, as a pre-signed WAV link.

Hits, loops and textures

Leave kind out and the server reads it off the words. One rule matters: keep tempo words out of a hit's prompt. "120 bpm" makes the model write a loop. duration_seconds overrides the default (up to 120 s), bpm and time_signature set the grid a loop is cut to, and seed makes a result reproducible.

One call, one sound

claude mcp add --transport http --scope user doseedo https://api.doseedo.com/mcp

doseedo_generate_sound {
  "prompt": "heavy wooden door slamming shut in a stone hallway",
  "kind": "hit"
}

Over HTTP it is the same job: POST https://api.doseedo.com/api/jobs with {"kind":"generate_sound","input":{…}} and an X-API-Key, then long-poll GET /api/jobs/{id}?wait=50. This is the finished job from October 22, 2026, trimmed. It took 20 seconds from submit to done:

{
  "kind": "generate_sound",
  "status": "succeeded",
  "output": {
    "files": [ "https://api.doseedo.com/api/files/…/stemphonic_327741615.wav" ],
    "duration": 2,
    "seed": 327741615,
    "sound_engine": { "engine": "stable-audio-open-1.0", "kind": "hit", "steps": 100, "gen_s": 17.19 }
  },
  "artifacts": [ { "url": "…/stemphonic_327741615.wav", "type": "audio", "expires_at": "2026-10-23T14:00:00.000Z" } ]
}

The file is a 2.0-second, 48 kHz stereo WAV, measured at −19.5 LUFS with a −0.5 dBFS peak. Download it with a plain curl -OJ <url> (no header, valid 24 hours), or hand the URL straight to the next tool.

Where the sound goes next

• Onto a picture. Place it at the hit's time with ffmpeg (adelay in milliseconds) and mix it under the music. The film-scoring guide has the whole command, ducking and loudness included.
• Onto a session. doseedo_add_audio_to_session puts the file on a DAW session, or doseedo_edit_session loads it into a Quick Sampler (load_quick_sampler_sample) so a MIDI note triggers it. With Dø Desktop the edit lands in your Logic project.
• Into a game or app. It's a WAV. Convert it with ffmpeg to whatever the engine wants.

The wrong tool for notes

A sound effect has no score. When the sound has pitches and rhythm that must be right, it is a part: write the notes and play them with doseedo_render_midi. That tool also covers electronic music from notes: instrument:"synth" plays your part on a patch Stable Audio Open makes for it (synth_role bass, pad, lead, pluck, 808), and instrument:"electronic" plays a drum part on a generated one-shot kit for a drum_genre. A whole song from a prompt is doseedo_generate_music.

What it costs

A sound is a render: 10 credits per 30-second window, so any hit, loop or texture up to 30 seconds is 10 credits. doseedo_quote answers before anything runs, with no sign-in. This is its answer for the door slam above, captured October 22, 2026:

{
  "object": "quote",
  "kind": "generate_sound",
  "estimate": { "generation": 10 },
  "basis": "10 credits per 30 s window (2 s → 1 window) → 10 credits"
}

Failed jobs are refunded, and the free tier covers the first sounds. Pricing has the plans.

Honest limits

• Generated, not recorded. Each sound is new and can surprise you, so listen before it ships and re-run with another seed if it misses.
• No voices. Stable Audio Open does not generate realistic voices, so it is no source for dialogue or sung lyrics.
• Prompts up to 300 characters, lengths up to 120 seconds, one sound per call.

Related

• Score a film with an AI agent
• Render MIDI with real instruments
• Music generation over MCP
• Put the sound on a track: editing sessions with Claude