video-use Review: What the Repository Proves About Its Model, Subtitles, Cost, and Data Path

Last updated: August 15, 2026

video-use is an MIT-licensed agent skill. A coding agent reads its instruction file, calls ElevenLabs to convert speech into timestamped text, writes an edit decision list, then drives ffmpeg. The repository ships instructions and six Python helpers. It ships no timeline interface and no packaged runtime.

Verdict in one sentence: this is an integration you take ownership of, not an editor you install, and the repository is unusually direct about handing you that ownership.

What you actually adopt

The repository contains an instruction file called SKILL.md and six helpers: transcribe.py, transcribe_batch.py, pack_transcripts.py, timeline_view.py, grade.py, and render.py. There is no timeline application and no hosted runtime inside it. Editorial judgment comes from whatever coding agent you point at the folder, and the setup prompt names Claude Code, Codex, Hermes, and Openclaw, so the tool is not tied to one vendor.

What the buyer supplies is longer than what the repository supplies: the machine, shell access for an autonomous agent, an agent account, an ElevenLabs key, an ffmpeg build, dependency management, and inspection of every render.

Decision artifacts are exposed rather than hidden. The workflow writes an edit decision list as JSON, session memory as Markdown, and cached transcripts as JSON. It also writes diagnostic PNGs, an SRT file, per-segment intermediate clips, and the final MP4. You can read why a cut was made. That does not by itself make the environment reproducible, and it does not establish that the pipeline completes.

One practical detail worth knowing before you budget: the helpers can be invoked directly with python helpers/<name>.py. The agent is required for the conversational layer that decides what to cut, not for the mechanical steps.

Sources: Browser Use, video-use repository, accessed 15 August 2026 · Browser Use, video-use SKILL.md, accessed 15 August 2026 · Browser Use, video-use install.md, accessed 15 August 2026

The model version is pinned in code

transcribe.py builds its request with model_id set to the literal string scribe_v1. The helper accepts flags for language and speaker count. It accepts none for the model. Changing it means editing the file.

The ElevenLabs API rate card prices Scribe v2 at $0.22 per hour and Scribe v2 Realtime at $0.39 per hour. On the version of that page reviewed here, there is no Scribe v1 line.

So the price most buyers will find while sizing this tool is quoted against a model the shipped helper does not request. That gap is not resolvable from the pages reviewed. Before committing volume, ask ElevenLabs in writing what a scribe_v1 request currently does and what it bills at, and treat the answer as a dependency question rather than a pricing footnote. A pinned model version is a maintenance obligation you inherit on install.

Two related facts, since both affect cost sizing. The helper requests speaker diarization on every call and caches each transcript by source file, so a re-edit does not re-bill. And ElevenLabs prices the same Scribe capability differently depending on whether you reach it through the API or the web interface, which we cover in full in our ElevenLabs pricing briefing.

Sources: Browser Use, video-use transcribe.py, accessed 15 August 2026 · ElevenLabs, ElevenAPI pricing, accessed 15 August 2026

Three risks in the subtitle path

Most coverage of this tool compresses everything below into one line about weak non-English support. The evidence separates into three layers with three different confidence levels, and a fix in one layer does not touch the others.

Cue segmentation. This part is verifiable in the code. render.py filters transcript entries to those of type word, takes them two at a time, joins them with a space, and uppercases the result. The cue length is fixed at two entries. ElevenLabs defines a word entry as a word in the language of the audio, and separately documents that its spacing entries do not apply to Japanese, Mandarin, Thai, Lao, Burmese, or Cantonese. Timestamped word entries therefore still exist in those languages. What the reviewed documentation does not establish is that two of them form a comparable caption unit across languages. The author of pull request 103 reports that Chinese entries arrived one per character and that the two-entry rule produced captions of separated characters with orphaned punctuation. That pull request has four commits, one automated bot review, and no human reply.

Font coverage. Also verifiable in the code. The subtitle style hardcodes FontName=Helvetica and passes it to ffmpeg through force_style, which overrides whatever the subtitle file specifies. The author of issue 118 reports that this produced empty boxes for a Chinese transcript on macOS, and that substituting a font with wider script coverage fixed it in their environment. A contributor has since responded that a fix is implemented in pull request 120.

Subtitle filter availability. The same reporter states that the ffmpeg installed by the command in the project’s own install guide lacked the subtitles filter entirely, so burn-in exited with an error while the SRT file itself was written correctly. That failure is language-independent and sits on the documented macOS path. It is a contributor report, not something reproduced here.

There is also a Windows report outside this path. The author of issue 125 states that helpers print arrow and dash glyphs to standard output, that Python on Windows falls back to the locale code page for a redirected stream, and that render.py therefore raises on its first progress line before extracting any segment. The same reporter states that changing the console code page does not help and that setting an environment variable does. The install guide provides macOS and Linux package commands and no native Windows path.

Sources: Browser Use, video-use render.py, accessed 15 August 2026 · ElevenLabs, Transcription documentation, accessed 15 August 2026 · GitHub, video-use pull request 103, opened 5 July 2026 · GitHub, video-use issue 118, August 2026 · GitHub, video-use issue 125, August 2026

What leaves the machine

transcribe.py extracts mono 16 kHz audio with ffmpeg and uploads it to the ElevenLabs speech-to-text endpoint. That is the design, not a configuration option.

ElevenLabs states that customer data is hosted and stored in the United States as standard, and describes data residency in the EU, India, and Singapore as an exclusive feature available to Enterprise customers. It states that Zero Retention Mode is available to Enterprise customers, applies to API traffic only, and is enabled by sending a specific logging parameter with each request. The reviewed helper does not send that parameter. For accounts without Zero Retention Mode, the documentation states that history preservation is enabled by default.

The audio path is the part that is fully documented. The visual path is not. The instruction file directs the coding agent to generate filmstrip and waveform PNGs with timeline_view and to read them at decision points. Those images contain video frames. Whether they leave your machine depends on which agent you run, how it authenticates, and how that provider handles image inputs. None of the reviewed material closes that question, so no one should tell you that video frames stay local.

For personal footage this may be an acceptable trade. For client interviews, recordings with identifiable employees, or anything under a confidentiality obligation, map four things before installing: what reaches ElevenLabs, what reaches your agent provider, where each stores it, and which retention controls your plan actually grants. Those are questions for your counsel and your contracts.

Sources: Browser Use, video-use transcribe.py, accessed 15 August 2026 · Browser Use, video-use SKILL.md, accessed 15 August 2026 · ElevenLabs, Data residency, accessed 15 August 2026 · ElevenLabs, Zero Retention Mode, accessed 15 August 2026

What free means here

The launch framing was “100% open source, 100% free”. The repository carries no license fee, which is accurate. The workflow it describes consumes two separately governed resources on top of that.

The agent layer is where a monthly estimate becomes unreliable, and the reason is authentication rather than price. Anthropic states that activity in Claude and Claude Code counts against the same usage limits, on a rolling five-hour session window with weekly limits on paid plans. Editing sessions read large transcripts and iterate, and that capacity comes out of the same pool as your other work.

What happens at the limit is a choice, not an automatic charge. Anthropic states that you can wait for the reset, change plan, or enable usage credits billed at standard API rates, and that all transitions to API credit usage require explicit user consent. There is one setting to check first: Anthropic states that if an ANTHROPIC_API_KEY environment variable is present, Claude Code authenticates with that key and incurs API charges instead of drawing on subscription usage. Many developers already have that variable set for unrelated reasons.

For a subscription-authenticated user with spare capacity, the immediate cash cost of one edit may be transcription only. An API key, an exhausted plan, or a different agent changes that route entirely. No reviewed document establishes token consumption, wall-clock time, failed-run waste, or correction time for a representative job, and those figures require a controlled run.

One boundary worth stating plainly. Browser Use publishes a pricing page, and it prices browser automation: sessions, proxies, and hosted agent tokens. Nothing on it prices video-use. The Browser Use Box page states that its preinstalled stack needs a Browser Use API key and your own Claude Code, so the hosted route adds a layer rather than removing one.

Sources: Anthropic, Choose a Claude plan, 20 May 2026 · Anthropic, Use Claude Code with your Pro or Max plan, 12 June 2026 · Browser Use, Pricing, accessed 15 August 2026 · Browser Use, Browser Use Box, accessed 15 August 2026

There is no pinned distribution

The documented install clones the main branch. On 15 August 2026 the releases page showed no published release and the tags view showed no tags. The repository tree contained no dependency lockfile, and pull request 108 proposes adding one.

The consequence is narrow and practical. Two installs on two dates are not designed to give you the same source and the same resolved dependencies. Nothing prevents you from making it reproducible: record the commit SHA, freeze the Python environment, and note the ffmpeg build. The project simply does not do it for you, and no upstream artifact exists to point at when something changes.

This says nothing about how responsive the maintainers are, and it does not establish that any reported defect still reproduces. It establishes that adopting this means owning a fork, and that the fork is your version control, not theirs.

Sources: GitHub, video-use releases, accessed 15 August 2026 · GitHub, video-use pull request 108, July 2026 · Browser Use, video-use install.md, accessed 15 August 2026

What this briefing cannot establish

The right column is longer than it looks. Closing it needs a paid account, one controlled run, and an invoice.

FAQ

Methodology

This is a document-first briefing. Future Stack Reviews holds no ElevenLabs account, has not installed video-use, sent no transcription request, ran no render, and inspected no invoice.

Sources reviewed on 15 August 2026: the video-use repository root, releases and tags views, contributors graph, SKILL.md, install.md, helpers/transcribe.py at commit fee55aa, helpers/render.py at commit 1200463, pull requests 103 and 108, issues 118 and 125, the ElevenLabs transcription documentation, API pricing page, data residency page and Zero Retention Mode page, the Anthropic plan and Claude Code support articles, and the Browser Use pricing and Box pages.

Code claims cite the file-level commits above rather than the branch name, because the documented install follows a moving branch and a link to it can change without notice. Claims sourced to a pull request or issue are attributed to the contributor who filed them and were not reproduced here. Repository counts and vendor prices are as displayed on the date given and are volatile.

Two absence statements appear, each bounded to a named page and date: no Scribe v1 line on the API rate card, and no native Windows path in the install guide. Neither asserts that no such thing exists elsewhere.

Verdict

Prototype video-use if your team can own a fork, pin its environment, and check every render by hand. Do not adopt it as supported production software on the strength of the repository pitch.

The repository exposes its edit plan as files and its helpers as short, readable Python, which is a real advantage over a closed timeline editor. What it does not supply is a release, a support boundary, a documented Windows path, a data-flow map, or any assurance about output. Those become your responsibility on the day you clone it, and that trade is the actual purchase decision.

Three things to settle before you commit. Ask ElevenLabs what a scribe_v1 request currently returns and bills. Run one subtitle-bearing render on your target operating system, in your target language, before planning any workflow around it. And check whether an ANTHROPIC_API_KEY is already set on the machine you intend to use, because that single variable decides which account gets billed.