Three free tools aim to give Claude video-editing agents a way to hear

Why the clip looks right but sounds wrong

As dev.to reported, a Thai-language post starts from a simple complaint. Claude can cut a video that looks good, yet the finished clip feels off to the ear. The post builds on a 75-second TikTok clip from the account @krit.thienn, posted on 27 Oct 2026.

The speaker in that clip says the trouble is timing. Sound is missing where it should be, or it doesn't match what is on screen.

The post's explanation is that an agent edits video by reading the code that builds the picture. It sees a timeline as numbers, such as a cut at 3.2 seconds, but has no idea what should be heard at that point. It adds that a sound placed on the timeline can end up 80 milliseconds behind the image.

Typical mistakes include a button-click sound laid over a shot where nothing is pressed. Another is music that swells just where it should drop out so speech stays clear.

Three tools, three layers of the problem

The post treats the tools as fixes for different layers: raw material, hearing, and placement.

Openverse is a searchable library of openly licensed images and audio. Its own API reports 5,223,435 audio items in total, made up of 3,986,828 from Wikimedia Commons, 644,709 from Jamendo and 591,898 from Freesound. An agent can query it without an account, though anonymous use is capped at 20 requests a minute and 200 a day. Each result comes back with license information.

Earshot describes itself as ears for a coding agent. It is open source under the MIT license and launched on 15 Oct 2026. When the author checked, it had 3,051 stars and 188 forks on GitHub, which fits the clip's claim of nearly three thousand stars in under two weeks.

Earshot turns sound into evidence an agent can reason about. It can describe a file in words, score how well the sound matches a description, and find the best stretch in a long recording. It can also say whether a sound was recorded for real or synthesized from code.

It runs on a CPU alone, so no graphics card is needed. It works as a command-line tool, a Python library, or an MCP server (a standard connector that lets an AI assistant use outside tools).

Making and reverse-engineering sounds

Earshot also has a command called forge. It lets an agent generate a sound from code and keep adjusting it until it sounds right. In version 0.4 the built-in library holds 252 sounds, 51 of them new in that release. They range from UI pops and wood knocks to bells and reversed cymbals.

A second command, fit, works the other way. It takes a recording and turns it into editable code. The author suggests this for cases where you like a sound from another clip and want your own version without using their file.

The author is careful about limits. Synthesized sounds don't suit everything. Natural sounds such as rain or a crowded room still need real recordings. And Earshot does not edit your video project: it produces audio files, and you still place them on the timeline in your own editor.

The tool the author could not verify

The third tool is called Reel SFX. According to the clip, it is a free Claude Code skill that finds every beat in a video, drops real recorded sounds onto the exact frame, and reads the editing project you are using. A skill is a set of instructions the agent follows, which means installing one gives it permission to act on your machine.

The speaker admits it ships with only 78 sounds, mostly interface sounds, so it needs Openverse for real sound effects.

The author searched GitHub and the web and found no public repository with that name. So everything about Reel SFX is the speaker's own account. The advice is to ask the clip's creator for the install source and read the code before running it.

The post is also open about its own origin. It says it was written by an AI model, run through an agent tool, under human control and quality checks. That is a reason to verify before you build a workflow on it.

The costs nobody invoices you for

Openverse is free and Earshot is open source, but the author points to three hidden costs. The first is license checking. Free in Openverse means no fee, not no obligation: some files use CC BY, which requires crediting the creator. If an agent picks files itself, have it log the license of each one so you don't hunt for sources at publishing time.

The second is processing time. Earshot analyzes audio on your own processor, so a clip several minutes long takes minutes, not seconds.

The third is judgment. An agent can tell you the sound doesn't match the picture, but it can't know what your audience likes. The final taste call stays with you.

A sensible order of work

The author's suggested sequence starts with listening: have the agent run Earshot's timeline command on the finished cut to see sync events and loudness. Next, search Openverse only for the sounds that are missing. Then generate any effects you can synthesize with Earshot's built-in library and render them as audio files.

The logic is that evidence comes before edits. Skip the listening step and the agent guesses from the picture alone. The author also warns against filling the whole clip with sound, saying over-filling is the most common cause of clips that tire the listener.

If you start with only one tool, the author picks Earshot, because it changes the agent from something that guesses at sound to something that can check it before delivery. The closing question is whether your work lacks sound, or lacks someone who hears it.