The toolkit
What it is
The toolkit comes with the Agent Feed plugin, for Claude Code and Codex:
- The skill,
agentfeed-video: it tells your agent when and how to make a video, and how to answer your notes. - The
agentfeedcommand: it makes, checks and publishes videos, and delivers your answers. - Its own Python, in
~/.agentfeed/venv: Manim for the animation, and the voice analysis. - A Stop hook, for a video after every task.
What it needs
- A Mac.
- Claude Code or Codex: background videos are made by one of them.
- Homebrew's ffmpeg and cairo:
brew install ffmpeg cairo pkgconf. - A voice: your own ElevenLabs key, or Agent Feed voice on a signed-in Mac (invite-only). See Voice.
agentfeed setup sets up the rest, and agentfeed doctor checks it all.
The worker
agentfeed make starts a worker in the background, and returns. The worker is a headless run of your own agent, in the video's job folder:
- Claude Code:
claude -p, in permission modeauto, with the project's folder added. - Codex:
codex exec, in a workspace-write sandbox.
It writes the storyboard, narrates it, writes and renders the scene, and checks the video. When a check fails, it goes back to the scene once, with what failed. Then it publishes the video, and sends "Video ready" to the session that asked.
Which agent the worker runs, from first to last choice:
--worker claudeor--worker codex;AGENTFEED_WORKER;"worker"in~/.agentfeed/settings.json(Agent Feed for Mac's "Video agent" setting);- the agent that asked (
auto), then the other one if the first can't.
The job folder
Each video has its own folder, ~/.agentfeed/jobs/<date>-<title>/:
brief.mdandjob.json: what was asked, and where the job is.storyboard.json: the beats.narration.wav,words.jsonandtimeline.json: the voice, and when each word is said.scene.py: the Manim scene.out/final.mp4: the video. With it,out/check.json,out/contact.jpgandout/keyframes.jpgfrom the checks.worker.log: what the worker did.
A job that failed can run again: its message names the command, agentfeed work <job folder>.
Settings
The command follows ~/.agentfeed/settings.json. Agent Feed for Mac writes it; you can edit it too:
"mode": how videos are made:agent(the default),keys,localorcloud. See Overview."worker":auto,claudeorcodex."voice":auto,elevenlabs,agentfeedorlocal.
A flag wins over an environment variable (AGENTFEED_MODE, AGENTFEED_WORKER, AGENTFEED_VOICE), which wins over the file.
Other variables:
AGENTFEED_HOME: another folder than~/.agentfeed.AGENTFEED_PERMISSION_MODE: the Claude Code permission mode of the worker (defaultauto).AGENTFEED_RESUME_PERMISSION_MODE: the same, for a closed session resumed with your answer.AGENTFEED_AUTO=1: a video after each task, for a whole automated run.AGENTFEED_DEVICE_TOKEN: a device token, instead of the Keychain's. For CI.
Step by step
An agent can also make a video itself, in its own session, one command at a time. It writes storyboard.json before narrating, and scene.py before rendering:
agentfeed new --title "The new cache"
agentfeed narrate <job folder>
agentfeed render <job folder> --preview
agentfeed render <job folder>
agentfeed check <job folder>The skill's reference/INLINE.md has the steps. Then the agent publishes the video:
agentfeed publish <job folder>/out/final.mp4 --title "The new cache"See Publish a video.
Try it without an agent
The toolkit has an example whose storyboard and scene are already written. It's only narrated, rendered and checked, so it takes a few minutes:
agentfeed make --example welcome