Tried browser-use/video-use: Coding Agents for Video Editing

I took a quick look at browser-use/video-use, an open-source project that lets coding agents operate video-editing workflows through natural-language instructions and code.

The concept is straightforward: describe an edit—such as trimming clips, adding captions, generating a highlight reel, or applying a repeatable transformation—and let an agent plan and execute the workflow. This is interesting because video editing becomes scriptable, testable, and easier to integrate into automation pipelines.

The repository gained +504 GitHub stars today, which suggests strong developer interest in treating video production as an agent-driven engineering problem rather than a purely visual task.

A practical gateway configuration could look like this:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_B_LOST_KEY",
    base_url="https://b-lost.com/v1",
)

response = client.chat.completions.create(
    model="claude-fable-5",
    messages=[
        {
            "role": "user",
            "content": (
                "Create a 30-second highlight video from ./clips, "
                "add subtitles, and export to ./out/highlight.mp4."
            ),
        }
    ],
)

print(response.choices[0].message.content)

The exact adapter names may change as video-use evolves, but the architecture is conventional: video-use as the agent layer, claude-fable-5 as the reasoning model, and an OpenAI-compatible relay as the transport layer.

For evaluation, I would track:

Metric	Why it matters
TTFT	Responsiveness during interactive editing
End-to-end render time	Practical workflow throughput
Edit success rate	Reliability across varied footage
Cost per 1M tokens	Budget predictability

B-Lost’s relay uses https://b-lost.com/v1, offers 20% off official list pricing, and supports standard OpenAI-compatible clients. For long system prompts and repeated editing instructions, its native Anthropic /v1/messages support and prompt caching could also reduce input costs, with a stated 90% discount on cache hits. The important caveat is to benchmark real workloads: model latency, tool execution, and video rendering may dominate total turnaround time.