ListenHub Hypit

Most ListenHub skills produce one artifact: an episode, a track, an image. listenhub-hypit produces a video project. It takes a reference clip, a brief, or your own footage, works out what the finished piece needs, generates every missing asset through ListenHub, and assembles them into a hypit project that renders to an MP4 on your machine — and that you can re-render after changing one line of dialogue.

The split is worth understanding before you start:

• ListenHub generates and transcribes. Images, video takes, speech, cloned voices, music, and word-level transcripts all come from ListenHub's public API, billed in credits.
• hypit is an external, locally installed CLI. It analyzes footage, holds the Script — hypit's structured record of every line, segment, and timing in the project — computes caption alignment, composites the result, and renders it. It runs on your machine and costs nothing to run.
• The bundled Provider is the bridge: it exposes two hypit capabilities — word-level alignment and background removal — backed by ListenHub instead of a third-party cloud.

No HypiHub account is required. The project, the render, and the keying all stay on your machine. The only data that leaves it is what the ListenHub API calls send — and the quote lists every one of them in advance.

Trigger

Invoke this skill with /listenhub-hypit, or use any of these phrases:

Requires ListenHub Skills to be installed — see Getting Started.

Quick Example

Remake this clip with my product in it, keep the original BGM, swap the voiceover to English:
https://example.com/reference.mp4

The agent probes the clip, plans the cut, and comes back with an itemised quote. Nothing is charged until you approve it.

What it chains

The skill does not reimplement generation. It routes each job to the ListenHub skill that already owns it, passing explicit models and parameters rather than inheriting that skill's defaults.

Prerequisites

The agent installs and runs everything itself — you are not expected to type these commands.

npx skills add hypit-ai/hypit -g
npm init -y
npm install @hypit/hypit@0.2.1
npm install --install-links <path-to-installed-skill>/listenhub-hypit/provider
npx hypit runtime init

<path-to-installed-skill> is wherever your agent installed ListenHub Skills — the agent resolves it itself, and the same directory holds the templates/ files referenced below.

Work in a dedicated video project directory and run every hypit command through npx hypit so the project's pinned version is used rather than whatever is on PATH.

The API key is stored through hypit's credential store, never written into the profile or a command line:

npx hypit auth login listenhub.local --slot apiKey --from /private/temporary/key-file
npx hypit auth status listenhub.local --json

Delete the temporary file afterwards. On Linux without a desktop keyring, hypit falls back to a user-private file; macOS and Windows use the system credential store.

Rendering needs a managed runtime (Chrome, OpenCV) that a fresh machine does not have:

npx hypit runtime up --runtime ./hypit.runtime.json
npx hypit doctor --runtime ./hypit.runtime.json --json

These local installs cost no credits. A MANAGED_PROGRAM_DOWN error is a runtime problem, not a failed generation — fix it before submitting anything billable.

How a production runs

npx hypit check narration.svrun --json
npx hypit plan narration.svrun --runtime ./hypit.runtime.json --json
npx hypit build narration.svrun --runtime ./hypit.runtime.json --follow

Four starting templates ship with the skill — narration, dialogue, action, and cutout, plus a shared styles.svs. They are examples to edit, not fixed formats: subject matter and character type are unconstrained, and a single timeline can mix spoken segments with pure-action ones.

Reusing the reference audio

Keeping sound you already have is the cheapest path through the whole workflow, and the skill defaults to it rather than regenerating.

You are responsible for holding the rights to any audio you did not create. The skill asks you to confirm this once; it does not verify the answer and does not clear rights on your behalf.

What it costs

Nothing billable runs before you see a quote. The skill lists every planned item — stills, takes per segment with model and duration, TTS character units, music tracks, transcription milliseconds, cutouts, clone saves — with the source of each price, the total, and your balance.

Prices come from the live estimate endpoints wherever one exists:

Video estimates come back in US dollars rather than credits, and the skill converts them at the service's current rate, keeping both figures in the quote. Transcription is 1 credit per started minute of audio, reserved rather than deducted until the task settles. A cutout is billed against Seedream's free image quota when the account still has some, but the estimate endpoint always reports the full credit cost — so the quote shows the worst case, and the reconciliation shows what was actually spent.

Items with no estimate endpoint are quoted from a dated rate table and marked as such. Existing assets, local extraction, cutting and mixing are listed at zero. Changing one line of dialogue or adding a variant re-quotes only the new work; a full re-confirmation is needed only when the account, scope, or budget changes.

Balances are recorded before and after each call, and every task's ID, model, and billed output is reconciled against the quote. Refunds for failures and unsettled reservations are listed separately.

The bundled Provider

The Provider adds exactly two hypit capabilities. Everything else still comes from the generation skills.

Transcription covers zh en ja ko vi th id ms fil hi ar fr de es pt ru it nl sv da fi no el pl cs hu ro bg hr sk. Requests in other languages are rejected at the planning stage rather than silently routed elsewhere.

Transcription is idempotent by content: the key is a SHA-256 of the WAV bytes, the language, and the Provider contract version, so re-running the same input returns the original task without reserving credits again — failed tasks included. Paying to re-transcribe is a deliberate act with a new key, never an automatic retry.

Limits

These are the skill's own boundaries, and it will not claim more than them:

• Reference fidelity is not guaranteed. How closely a generated take reproduces a reference clip's motion, timbre, and lip sync varies with the model, the prompt, and the source material, and cannot be predicted before you render. Check the result; no reference video is promised an exact reproduction.
• Stem separation is lossy. Pulling vocals or BGM out of a mixed track leaves artifacts. Listen to the result before building on it — no mixed track separates cleanly.
• Cutout is a re-render, not a matte. The green-screen pass can redraw detail, and it renders at the nearest supported frame size rather than the source's exact dimensions — read the output's real dimensions before writing them into the project. Subjects wearing green, or made of glass or anything translucent, are the wrong input for this path. Video green-screen work goes through ffmpeg, not this capability.
• Reference audio is not passthrough. Driving a generation with the original voice produces a new performance; check the dialogue, pacing, and timbre of what comes back.
• TTS lines must match the Script exactly. Word-level alignment fails when the spoken text and the Script's text diverge.

This skill and its Provider are MIT-licensed. hypit is an external dependency under its own licence (hypit-ai/hypit), which requires you to keep the hypit CLI, Studio, run report, and manifest branding and copyright intact, and not to redistribute hypit as a hosted or paid product.

Related