HeyGen + n8n: script to talking avatar without living in the GUI

Use this stack as a page: the studio avatar pipeline.

The AI avatar generator card is the only one that names the pair on the art: HeyGen + n8n. The pipeline is printed in order: script → generate avatar → add voice → export. That is the whole product. The “fully automated / ready for social” bar is marketing around a render job.

A talking-head avatar is not faceless. If you wanted no person on screen, use the faceless post. If you wanted a real camera, this post is the wrong spend.

The stack:

1. Script — Claude or ChatGPT.
2. Avatar render — HeyGen. Swap Synthesia if that is already procurement.
3. Voice — HeyGen’s own voices, or ElevenLabs / Murf AI if you need a library you reuse outside HeyGen.
4. Orchestrate + export — n8n. Zapier if you do not self-host.

This is not our avatar studio. It is the handoff so you do not paste a script into ChatGPT and hope a .mp4 appears.

Why the GUI does not scale and a chatbot cannot replace HeyGen

HeyGen’s job is a digital human that lip-syncs. Claude cannot do that. Sora and Kling AI generate scenes; they do not give you a consented, reusable presenter with a stable face across 40 training videos. Different row in the matrix, different contract.

HeyGen has a free row and paid from $24/month on our matrix. Localization — the same avatar, more languages — is the reason companies pay. The card’s “multiple languages” is that feature, not a chatbot translating a YouTube description.

Synthesia is the enterprise cousin: no free row, paid from $30/month. Training, HR, localized explainers. If security review already picked Synthesia, do not add HeyGen “because the infographic said so.”

n8n is free to self-host, cloud about $22/month. It should call HeyGen’s API (or a watched folder) and then park the file in Drive/S3 and ping Slack. Humans still approve the first frame of a new avatar. An unattended clone of a real employee is an HR incident.

The four jobs (and what each must not do)

Claude owns the script, not the face

Claude $20/month paid: 60–90 seconds, pronunciation notes for product names, a pause list, language code. ChatGPT $20/month paid if that is the house LLM.

Do not generate a new avatar description every run. The avatar is an asset. The script is the variable.

Keep the tone in the same family as Claude for essays if the video is explanatory. Keep it shorter than an essay. Avatars look drunk when the line is a paragraph.

HeyGen owns the human

Stock avatar vs avatar you have rights to. Lip-sync. Optional gestures. Export 1080p or whatever the social crop needs.

Do not upload a celebrity. Do not use a coworker’s face without a written yes. The same line we use for ElevenLabs cloning applies to faces.

Voice owns identity across tools

If every video stays in HeyGen, their voices are enough. If the same person also narrates Kling B-roll, put the voice in ElevenLabs (free, paid from $5/month) and keep HeyGen on picture. Murf AI is paid from $23/month, no free row — another studio voice + avatar option, not a second HeyGen.

Pick one voice id. Changing narrator every episode is how training videos feel like a demo reel.

n8n owns Run Automation

Webhook or cron → fetch next script from a sheet → POST HeyGen → wait (polling, not a tight loop) → download → rename → notify. Zapier $19.99/month Professional if the team already lives there.

Do not fire twenty renders because the sheet had twenty draft rows. Status column: draft | approved | rendered. Only approved hits the API. That is the difference between automation and a credit fire.

One video, six steps

1. Claude writes the script into the sheet.
2. Human marks approved (and language).
3. n8n sends script + avatar id + voice id to HeyGen.
4. Poll until ready. Save mp4.
5. Optional: burn captions, or export a subtitle file.
6. Human posts, or a separate publish Zap — do not hide publish inside the render flow until the avatar is trusted.

Feature bar on the card: realistic avatars, multiple languages, fully automated, ready for social. Translate: HeyGen, HeyGen, n8n, and a human the first month.

What this stack is not

It is not Sora UGC. Different camera.

It is not a replacement for Synthesia if that is already the standard. Compare them on HeyGen and Synthesia tool pages, not in a prompt.

It is not unlimited localization quality. A native speaker still checks the German take.

The page with the four matrix slots is the studio avatar pipeline. HeyGen owns the face; ElevenLabs owns the voice library; Descript masters. n8n can still be the clock if you already live there.