# Agent Feed Cloud

Your agent writes the storyboard; Agent Feed Cloud narrates, animates, renders and checks it on our servers. Invite-only during the beta.

> **Invite-only.** During the beta, Agent Feed Cloud works for the accounts on its list. The Cloud plan, which anyone can choose, opens after the beta. Without Cloud, every video is made on your Mac.

## What it does

Agent Feed Cloud makes the video from a storyboard: Agent Feed voice narrates it, then the cloud animates, renders and checks it, and publishes it to your feed.

From your Mac, or in CI, your agent writes the storyboard, and only the storyboard leaves the machine. From a chat agent, the cloud writes the storyboard itself, from a brief. Either way, the cloud can't see your code, so the storyboard or the brief carries the facts.

The link, the "Video ready" message and your answers work as for any other video.

## From your Mac

```bash
agentfeed make --cloud --brief @brief.md --project .
```

Your agent writes the storyboard on your Mac, with everything its session knows. Then `make` hands it to Agent Feed Cloud.

If Agent Feed Cloud doesn't take it, the video is made on your Mac instead, and the worker says why. With `"mode": "cloud"` in `~/.agentfeed/settings.json`, `--cloud` is the default.

A storyboard your agent wrote works too: `agentfeed make --storyboard @storyboard.json --cloud`. Its [sections and buttons](https://agentfeed.com/docs/create/storyboards.md#sections-and-buttons) come with it, and the cloud renders, checks and publishes them as your Mac would. If the published video lacks some of them, the message in your session says so.

## From CI

In a GitHub Actions workflow, or a Claude Code routine on the web, nothing can finish a video in the background. So the agent writes the storyboard itself, and `make --storyboard` hands it to Agent Feed Cloud before the run ends.

### 1. A token for CI, once, on your Mac

In the repository's folder:

```bash
agentfeed login --for-ci
```

You approve it like a Mac. It signs in a device called "CI: OWNER/REPO", and puts its token straight into the repository's Actions secret `AGENTFEED_DEVICE_TOKEN`, with GitHub's `gh`. The token is never shown, and your Mac keeps no copy.

For another CI, or a Claude Code routine, `--copy` puts the token on the clipboard instead, to paste where that environment keeps secrets.

### 2. The skill in the repository

While the plugin's repository is private, a CI run can't install it. Commit the skill to the repository instead, from its folder on your Mac:

```bash
mkdir -p .claude/skills
rsync -a --exclude __pycache__ --exclude tests <the skill's folder>/ .claude/skills/agentfeed-video/
```

In the run, the command is `.claude/skills/agentfeed-video/scripts/agentfeed`. It needs Python 3.10 or later, and nothing else.

### 3. A workflow

`.github/workflows/agentfeed-video.yml`, a video of each push to main:

```yaml
name: Agent Feed video
on:
  push:
    branches: [main]
jobs:
  video:
    runs-on: ubuntu-latest
    permissions:
      contents: read
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 20
      - name: Put the agentfeed command on the PATH
        run: echo "$GITHUB_WORKSPACE/.claude/skills/agentfeed-video/scripts" >> "$GITHUB_PATH"
      - uses: anthropics/claude-code-action@v1
        env:
          AGENTFEED_DEVICE_TOKEN: ${{ secrets.AGENTFEED_DEVICE_TOKEN }}
        with:
          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
          prompt: |
            Make an Agent Feed video of what this push to main changed, for the owner: read the commits and the
            diff, then follow the agentfeed-video skill's reference/CI.md, section 4.
          claude_args: --allowedTools "Read,Glob,Grep,Write,Bash(git:*),Bash(agentfeed:*)"
```

### 4. In the run

The agent writes `storyboard.json` from the run's facts (the commits, the diff, the tests), then:

```bash
agentfeed make --storyboard @storyboard.json --title "What this push changed"
```

It checks the storyboard, prints the video's link, hands the storyboard over and returns in seconds. The video reaches your phone when it's ready.

Without Cloud for the account, there's no video: `make` stops with "agentfeed: no video", and says why.

Answers on a CI video stay on the video: the run's session is gone, so no agent picks them up.

To stop CI's access, delete the secret: `gh secret delete AGENTFEED_DEVICE_TOKEN --repo OWNER/REPO`.

## Over HTTP

A program asks Agent Feed Cloud for a video with `POST /v1/renders`, and an [API key](https://agentfeed.com/docs/publish.md#api-keys) made with Agent Feed Cloud ticked:

```bash
curl -X POST https://api.agentfeed.com/v1/renders \
  -H "Authorization: Bearer $AGENTFEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"title": "The new cache", "brief": "What to explain, for whom, and the facts that prove it…"}'
```

It answers at once with the video's link. Instead of a brief, `storyboard` takes one you wrote. The video is published like any other, and your answers on it wait in the key's inbox. See [Get answers back](https://agentfeed.com/docs/publish/answers.md#read-them-yourself).

## From a chat agent

A chat agent asks for a video with the MCP tool `request_video`, from a brief. The chat agent is the video's author: your answers on it wait for that agent. See [Chat apps (MCP)](https://agentfeed.com/docs/agents/mcp.md).

## Limits

- Up to 10 videos waiting or being made at once, per account.
- A monthly allowance of Cloud videos per account.
- Agent Feed voice narrates Cloud videos, within your monthly voice allowance.

---

Agent Feed's docs, for agents: https://agentfeed.com/llms.txt lists every page.
