The night shift for your stack
24/7 AI video editor: wake up with your video already edited
Approve a script before bed. Overnight it records the voice, cuts the video and checks every step against your list. In the morning the draft is waiting for your OK.
- 11:04 PMScript approved by you
- Voice recorded, every line checked
- Pictures logged with their source
- Video cut, captions and sound checked
- Report written, with the cost of each paid step
- 6:30 AMDraft ready. Waiting for your OK
AI Stack Library · Content and media
A 24/7 AI video editor is a chain of steps, from an approved script to a finished cut, that an AI agent runs on an always-on machine without anyone at the keyboard; it is worth having only if every step ends in a check, and the agent stops and waits for a person before anything is published.
| If you are | Start with |
|---|---|
| A business posting one or two videos a week | A dedicated always-on machine, one agent running each step in order, and a person who watches every draft before it goes out |
| A business posting every day across several channels | A dedicated always-on machine, a separate folder and checklist for each channel, and a spending cap on every paid step |
| A business building video into a product it sells | A dedicated always-on machine and an agent SDK, such as the Claude Agent SDK, so your own code controls each step and its checks |
What runs while you sleep
The agent works from a folder: the approved script, the voice track, the pictures and clips, and a checklist. It cuts the video with a tool such as FFmpeg, checks the result, and leaves a draft and a short report for the morning.
Several tools can run an agent with nobody at the keyboard. Claude Code has a non-interactive mode, started with claude -p, that exits with code 0 when a run succeeds and a different code when it fails, so a script can tell a good night from a bad one. Goose and the Claude Agent SDK are other options.
The steps, in order
- A person approves the script. The agent never writes a fact into the script that nobody checked.
- Voice: the agent records the narration and checks that every line of the script was spoken, in order.
- Pictures: every image and clip comes with a note of where it came from and whether you may use it.
- Cut: the agent assembles the video and checks its length, sound level and captions against your checklist.
- Report: the agent writes what it did, what each paid step cost, and anything it could not finish.
- A person watches the draft and decides: publish, change, or drop.
Why the checks matter more than the editing
Cutting video is the easy part for an agent. The hard part is the line that was skipped, the clip nobody has the rights to, and the paid step that ran forty times instead of once.
So every step ends in a check the agent cannot talk its way past, and a failed check stops the line. A draft that waits for you is better than a wrong video that went out at 3 a.m.
Settings to get right before the first night
- Unattended means nobody can answer a question. In Claude Code, the --permission-prompts none flag denies anything that would have asked a person, and the run carries on without it.
- Put a budget on every run. Claude Code's --max-budget-usd flag stops a run at the cap you set.
- Keep the cost of each run. With --output-format json, Claude Code reports an estimated cost for the run alongside the result.
- Give each channel its own folder, so one channel's files never end up in another channel's video.
Questions
Can the agent publish the video itself?
It can be set up to, but we would not start there. Let it prepare the file, the title and the description, and have a person press publish until the drafts have earned your trust.
Why a dedicated machine instead of a laptop?
A laptop sleeps, travels and gets closed. An editor that works around the clock needs a machine that stays on, stays plugged in and does nothing else.
Does it need an expensive computer?
Not for cutting. FFmpeg runs on ordinary hardware. The paid parts, such as voice and generated clips, usually run on the vendor's servers, so the cap on each paid step matters more than the machine.
Sources
- Claude Code: Run Claude Code programmatically, checked 2026-10-22
- Claude Agent SDK overview, checked 2026-10-22
- FFmpeg, checked 2026-10-22
- Goose, checked 2026-10-22