video-to-sfx

Sonilo Video-to-SFX

Hand Sonilo a video and it generates sound effects matching what it sees — footsteps, impacts, ambience, UI sounds, whatever the scene calls for — or pin specific sounds to specific moments with segments. Generation runs as an async task on the backend; the tools poll internally and hand back the saved file.

Setup: See the setup-api-key skill.

⚠️ Cost: every tool below makes an API call that may incur charges. Only call it when explicitly requested.

Transport: MCP or CLI

Pick one at the start of the session and stay on it. Do not mix the two inside a single job, and do not announce the choice.

1. Sonilo MCP tools visible in this session (video_to_sfx and friends) — use them. This is the preferred path: it needs no shell, and it is the only one that survives a very long generation. If a call fails to authenticate — rather than failing on its inputs — this transport is not usable in this session: go to 2 instead of retrying it.
2. No usable Sonilo MCP tools, but sonilo account exits 0 — use the CLI commands below. Same API, same account, same credential file. Probe with sonilo account, not sonilo whoami: whoami exits 0 even when signed out, so it cannot tell the two states apart.
3. Neither — stop and run the setup-api-key skill. Do not call api.sonilo.com with curl to work around it; both transports handle uploads, polling and retries that a bare request does not.

Quick Start

MCP tool call (recommended)

video_to_sfx(
    video_path="~/Desktop/action-scene.mp4",
    prompt="Footsteps on gravel, distant traffic, a door slam"
)

Python (pip install sonilo)

from sonilo import Sonilo
 
client = Sonilo()  

JavaScript / TypeScript (npm install sonilo)

CLI (npm install -g sonilo-cli or pip install sonilo-cli)

Always async under the hood — the CLI submits and polls for you. --format accepts wav|mp3|aac|flac.

cURL (raw REST API, no MCP host)

Every call is task-based: the endpoint returns {"task_id": ...} (HTTP 202), and the result is fetched from GET /v1/tasks/{task_id} once status is terminal. The MCP tools do this polling for you and return the saved path directly — you only see the task_id if the call times out (see task-recovery).

Tools

Parameters

segments rules

Validated by the backend before any charge — an invalid list is rejected with a 422/400 and nothing is billed:

• First segment's start must be 0.
• Segments must be contiguous: each end must equal the next segment's start.
• Every end must be greater than its start.
• Every prompt must be non-empty, max 200 chars.
• The last end must not exceed the video's actual duration.
• Max 30 segments total.

Prompting

No prompt is required — the model reads the cut. Quality comes from a time-segmented action map: what is on screen, what it's made of, what it does, second by second. The footage is the source of truth.

Before a paid call: probe the exact duration and existing audio, respect the 180 s cap (over = 422 reject, never truncated), and get sign-off — failed runs auto-refund, but your own retry is a new charge.

• Full pre-flight (inspect the video, caps, credits, verification): references/preflight.md
• Action-map craft (scene bed, sound bundles, materials vocabulary, segment rules): references/sfx-prompting.md

Workflow Tips

• Leave prompt/segments unset to let Sonilo read the whole video and decide; use segments when you need specific sounds pinned to specific moments (e.g. a punch landing at 2.3s, a door slam at 5.0s).
• Want the video back with SFX baked in? Use video_to_video_sfx instead of video_to_sfx.
• Prompting: be specific and combine elements — "Heavy rain on a tin roof" beats "Rain".
• Don't confuse this with music. For a background score or soundtrack, use video-to-music instead. To generate both music and SFX together in one balanced, single-charge call, use video-to-sound.
• No footage? text-to-sfx generates a single clip from a description alone.
• Don't know what it should sound like? Run video-analysis first: one call returns a section plan plus ready-to-use generation prompts read off the footage, which beats guessing a prompt and rerolling. It is a paid call that generates nothing, so use it when the brief is genuinely unclear — not when the user already told you what they want.

Recovering a Timed-Out Call

Every tool here is async on the backend already; a long generation can still exceed TIME_OUT_SECONDS. If it does, the error carries a task_id — the job keeps running (and is already charged). Call get_sfx_task(task_id) — get_generation_task(task_id) on the hosted server — later to retrieve the result; see task-recovery.

Output Files

• video_to_sfx: saved in the requested audio_format (.wav/.mp3/.flac, or .m4a for the aac default), named from the prompt (slugified) or sfx-<first 8 chars of the task id>.
• video_to_video_sfx: a single .mp4 with the SFX muxed in.

Error Handling

Common errors: 401 invalid key, 402 insufficient balance / trial exhausted, 413 file too large, 422 invalid parameters or malformed segments, 429 rate limit. See the account skill.