ossclip.

A local-first CLI that turns a talking-head take into a finished short: filler words and dead air cut, word-timed kinetic captions, face-aware framing, and LLM-planned, code-rendered on-screen graphics — title cards, stat cards, diagrams, terminal and chat mockups — planned against a hand-built, Zod-typed scene library. Plus a cover image sized for the platform.

Your footage never leaves your machine. Transcription is local (whisper.cpp), rendering is local (Remotion); the only network calls are the LLM planning ones — on your own API key, or on your existing Claude Code subscription.

Scope, honestly: ossclip is at its best polishing a take you have already cut down. For long-form input, --clip <seconds> selects the single strongest window (the producer's judgement, sentence-snapped) and produces only that — one clip, not N. Without it, 20 minutes in is a polished 20 minutes out.

AI can make mistakes: the cut, captions and graphics are generated — review the output before publishing.

Left: the raw take. Right: what ossclip produce returns.

One ossclip produce run, end to end — every stage says what it did and why.

Install

Two commands, on macOS, Linux, or Windows (plain PowerShell — no WSL, no admin rights). The only prerequisite is Node ≥ 22 from nodejs.org.

npm install -g ossclip
ossclip setup

ossclip setup provisions everything into one folder (~/.ossclip): a static ffmpeg build, a prebuilt whisper.cpp whisper-cli (on macOS both come via Homebrew), and the transcription model — small.en, ~466 MB, the biggest piece of a ~600 MB total. It shows the plan with sizes and asks before downloading, resumes interrupted downloads, verifies checksums, and skips anything you already have — an ffmpeg already on your PATH stays yours. Paths are recorded in ~/.ossclip/config.json, so nothing edits your PATH. Uninstall = npm rm -g ossclip plus deleting ~/.ossclip.

Setup also offers to save an LLM key for --produce: a logged-in Claude Code is detected automatically, or paste an ANTHROPIC_API_KEY or GEMINI_API_KEY. Skip it freely — cut + captions run fully local without one. If anything looks wrong later, ossclip doctor prints a line per prerequisite and the exact fix.

Licence note, since setup downloads binaries: the static ffmpeg builds it fetches (BtbN) are GPL, downloaded onto your machine at your request — nothing GPL ships inside the MIT npm package.

Manual install (if you'd rather own the toolchain)

brew install ffmpeg whisper-cpp        # macOS; Linux/Windows: ffmpeg from your package manager,
                                       # whisper-cli from https://github.com/ggml-org/whisper.cpp/releases

mkdir -p ~/.ossclip/models
curl -L -o ~/.ossclip/models/ggml-small.en.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin

ossclip doctor checks all of it — with the per-platform command for anything missing. Binaries can live anywhere: point OSSCLIP_FFMPEG / OSSCLIP_WHISPER (or config.json) at them. Setup and manual install compose — setup only ever fills the gaps doctor would flag.

small.en is the default. A mistranscribed word ends up in your captions and on a graphic, so accuracy beats speed here — and --produce runs a repair pass over the transcript before anything is drawn. ossclip setup --model <name> downloads base.en (~142 MB) or medium.en (~1.5 GB) instead.

Quick start

# The whole thing: cut + captions + LLM-planned graphics + cover
ossclip produce input.mp4 --produce -o out.mp4

# Just the cut and captions — no LLM, no network
ossclip produce input.mp4 -o out.mp4

# See what would be cut and why, without rendering
ossclip produce input.mp4 --no-render

# Long-form in, one short out: the strongest ~60s window
ossclip produce podcast.mp4 --produce --clip 60 -o clip.mp4

# Check the whole toolchain first
ossclip doctor

# Open the editor over what a run produced (bare `edit` opens a project picker)
ossclip edit "<work directory>"

# Landscape (YouTube/desktop) instead of the 9:16 default
ossclip produce input.mp4 --produce --aspect 16:9 -o out.mp4

How it works

• Transcribe & repair — whisper.cpp locally, then an LLM pass that fixes mishearings (cloud code → Claude Code) under a strict gate: a rewrite is refused, a mishearing is fixed, and every repair is logged with what was heard.
• Cut — silences and fillers go by measured thresholds (--cleanup exact…aggressive); the report says what was cut and why.
• Captions — word-timed kinetic captions with an active-word highlight, routed around the platform's safe areas, the active graphic, and any text burned into the source.
• Producer — the LLM plans a beat sheet and scenes against the component library, steered by a measured framing brief (a wide band never lands on a close-up), grounded against the transcript so a stat it prints is a stat that was said.
• Verify — layouts that would crop the head are repaired, graphics route around the source's own on-screen text, on-screen copy is reconciled with the captions.
• Render — Remotion, locally, plus a cover JPEG at the OUTPUT's aspect: 9:16 framed for the profile grid's square crop, 16:9 framed as a thumbnail shown whole.

The work directory

Every run writes <input dir>/.ossclip/<name>-<hash>/ — the hash is of the source's CONTENT, so the same footage reuses its cache (rename it and you still land in the same place; --workdir starts a separate project, deleting the directory forces a clean run): the transcript, the analysis, production.json, render-props.json, report.txt, usage.json, the cached LLM plan, and command.json — the recorded invocation the editor's Render button replays. It is a cache: delete it to force a clean run; keep it and re-runs are near-instant.

Each run appends to usage.json — one entry per run with its provider, models, tokens and cost — and stamps the same provenance onto production.json, so a workdir always says who planned it even when a later run answers entirely from cache.

Models & cost

Without --llm, ossclip picks: GEMINI_API_KEY → ANTHROPIC_API_KEY → the Claude Code CLI (your Pro/Max subscription — plan usage, not API credits). Calls are tiered: editorial judgement (the beat sheet, transcript repair) goes to the main model; mechanical calls go to a smaller sibling.

▸ llm: 3 calls · 130,749 in / 8,321 out tokens · ~$0.72 of API-rate work, covered by the subscription · 92s

Reported token counts are exact where the provider reports them and marked (est) where derived; a model with no known price gets tokens and no cost guess. Per-call breakdown in report.txt, raw records in usage.json.

Editing

Select, drag, retype — every edit lands in overrides.json, which the producer never overwrites.

ossclip edit "<work directory>"     # opens http://127.0.0.1:5174
ossclip edit                        # bare: pick from recent projects, or browse

Keybinds

Press ? in the editor for this list in-app.

transport

selection

editing

view

Scene components

Every component sizes its type to the slot it is given (nothing overflows the platform-safe area), animates in with a staggered rise, and leaves through a uniform exit fade. All text on them is editable in place.

Layouts

Layout places the video; the component always renders. Portrait (9:16) and landscape (16:9) each get frame-appropriate geometry — the split axis follows the frame's long edge.

produce flags

Configuration

~/.ossclip/config.json — all optional:

{
  "model": "small.en",
  "speaker": "Ahsan, host of the Code with Ahsan channel",
  "fastModel": "claude-haiku-4-5-20251001",
  "ffmpegPath": "ffmpeg",
  "whisperPath": "whisper-cli",
  "modelDir": "~/.ossclip/models",
  "browserExecutable": "/path/to/chrome",
  "pricing": { "claude-opus-5": { "inputPerMTok": 15, "outputPerMTok": 75 } }
}

Provider keys come from the environment, and ossclip loads .env files before choosing a provider — first hit wins per key, a real environment variable always beats a file: $OSSCLIP_ENV_FILE → .env walking up from the cwd → ~/.ossclip/.env.

Env vars override the file: OSSCLIP_FFMPEG, OSSCLIP_FFPROBE, OSSCLIP_WHISPER, OSSCLIP_MODEL_DIR, OSSCLIP_MODEL, OSSCLIP_FAST_MODEL, OSSCLIP_SPEAKER, OSSCLIP_BROWSER.

Licensing

ossclip's own code is MIT — see the repo's LICENSE.

Rendering uses Remotion, which is source-available under its own two-tier licence, not MIT: free for individuals, non-profits, and for-profit companies up to the size stated in its terms; larger for-profit companies need a paid Remotion Company Licence. If you use ossclip inside a company, check Remotion's LICENSE — its terms are authoritative.