Source: ai-research/heygen-com-skills-bundle-2026-05-14.md — README from heygen-com/skills GitHub repo (MIT, 232 stars, v3.1.0 released 2026-04-27). Refreshed 2026-09-29 with raw/Como_Automatizar_Videos_con_IA_-_Claude_+_HeyGen_Paso_a_Paso_Tutorial_2026.md (HeyGen Spain community webinar, 2026-08-19, presenter Anastasia Martiniuk, HeyGen ambassador; YouTube 9oGeWHrSN_w; Spanish, notes translated).

Repo: https://github.com/heygen-com/skills Stars: 232 Language: Shell (100%) License: MIT Version: v3.1.0 (2026-04-27, 8 releases total)

HeyGen’s official, vendor-published three-skill bundle for AI agents to drive avatar video creation end-to-end: heygen-avatar (persistent digital twins from photos + voice synthesis), heygen-video (idea → scripted-and-delivered video), heygen-translate (175+ language localization with voice clone + lip sync). Sibling to skills in the vendor-published-agent-skills category — both ship Markdown-based skills installable via gh skill install, both target Claude Code / Cursor / Codex / OpenClaw and 7+ other runtimes, both function as the CLI+Skill alternative to that vendor’s MCP. Distinct from hyperframes (HeyGen’s composition framework — HTML-to-MP4 rendering, separate repo). The skills bundle is the vendor surface for HeyGen’s three production capabilities (avatar / video / translate); Hyperframes is the composition surface for stitching motion graphics, captions, and transitions around that output.

Key Takeaways

  • Three skills, one repo. heygen-avatar + heygen-video + heygen-translate, all under the same MIT-licensed heygen-com/skills repository at v3.1.0. Install one or all three independently via gh skill install heygen-com/skills <skill-name>.
  • heygen-avatar = persistent digital twins from photos. Trains a reusable avatar identity (the same productized surface as HeyGen Avatar V) and bundles voice synthesis so the avatar can speak in dialogues across multiple videos without re-training.
  • heygen-video = idea-to-scripted-video. Operator describes the target output (intro, outreach, explainer, update), skill returns a scripted video with avatar delivery + style recommendations. Productizes the “describe what you want, get a draft video” surface that previously required hand-crafted prompts in HeyGen Studio Automation.
  • heygen-translate = 175+ language localization. Takes a finished video, applies voice cloning, and re-lip-syncs the avatar in the target language. Same 175+ language coverage cited in Avatar V’s technical report; now exposed directly as an agent-callable skill.
  • Runtime spread is the widest we’ve tracked. Eleven runtimes named in the README: Claude Code, Cursor, Codex, OpenClaw, Gemini CLI, Copilot, Junie, Goose, OpenHands, Amp, Cline. Same gh skill install command, same SKILL.md contract, same auth flow.
  • Three install paths. Recommended: gh skill install heygen-com/skills <name> per skill. Alternative: ClawHub one-shot clawhub install heygen-skills. OpenClaw users: openclaw plugins install clawhub:@heygen/openclaw-plugin-heygen. Power users: clone the repo into the agent’s skills directory directly.
  • Dual-auth: API key OR ChatGPT-style OAuth via HeyGen plan. Operators with existing HeyGen subscriptions can run these skills via MCP/OAuth against their plan credits (no separate API-key budget). Operators without a HeyGen plan pay direct via API key. Same unit-economics pattern as the Hermes Codex App-Server Runtime (subscription-not-API-key) and the ChatGPT Ads access path (managed-tier credits vs self-serve).
  • Composable with Hyperframes. Operators can chain the three skills with the Hyperframes HTML composition framework: heygen-avatar trains the identity, heygen-video produces the talking head, Hyperframes composes captions / lower-thirds / b-roll / transitions around it, heygen-translate localizes the final cut. Two HeyGen repos, full-pipeline coverage.
  • Sister to the Higgsfield skills bundle pattern. Both vendors converged on the same surface in the same April-May 2026 window: vendor-curated skills bundle that wraps the vendor’s existing API/MCP/CLI in a Markdown-based skill format the agent runtime loads at session start. Pattern signal — expect Anthropic-blessed marketing-AI vendors (e.g., Adaline, TinyFish, Higgsfield, HeyGen) and Anthropic-adjacent dev-tool vendors (e.g., Vercel, Linear, Cloudflare) to ship analogous skills bundles over the next quarter.

What each skill does

heygen-avatar (persistent digital twins)

Trains a reusable avatar identity from a photo (or photos) + voice characteristics. Persists the identity so subsequent video generations reuse the same look, voice, and gestures without re-training. Same value prop as Avatar V’s video-reference conditioning, now exposed as an agent skill — the skill encapsulates the upload, training, and re-use steps so the agent invokes them via natural language rather than via direct API calls.

Use case: spin up a digital twin once per brand / personality / fictional character, then reference that twin across the rest of the workflow.

heygen-video (idea-to-scripted-video)

Takes a brief from the operator (“a 30-second cinematic intro,” “a video update for my team,” “an outreach video for potential investors”) and returns a scripted video with avatar delivery + style recommendations. The skill handles script generation, voice synthesis, avatar choice (or reuse of a previously-trained heygen-avatar identity), pacing, and delivery style. Closes the “describe what I want, get a draft” gap that operators previously crossed by hand-prompting Claude / GPT to write the script first, then manually feeding the script into HeyGen Studio.

Use case: any cold-start video brief where the operator doesn’t already have a script — internal updates, marketing intros, sales-enablement videos, briefings.

heygen-translate (175+ language localization)

Localizes a finished video into another language with voice cloning (preserves the speaker’s vocal identity in the new language) and lip-sync (re-renders the avatar’s mouth movements to match the new audio). Same 175+ language coverage as Avatar V. Skill encapsulates the upload-to-localize flow so the agent invokes it with a target language string.

Use case: any agency / DTC / SaaS workflow where one finished asset needs to ship in multiple language markets — turns localization from a separate production pass into a one-line agent invocation.

Install paths

PathCommandWhen to use
gh skill install (recommended)gh skill install heygen-com/skills heygen-avatar (one per skill)Default for Claude Code / Cursor / Codex / Gemini CLI / Copilot / Junie / Goose / OpenHands / Amp / Cline. Pick which skills you actually need; skip the rest.
ClawHub one-shotclawhub install heygen-skillsAll three skills at once. ClawHub manages the install + auth wiring.
OpenClaw pluginopenclaw plugins install clawhub:@heygen/openclaw-plugin-heygenOpenClaw / Hermes deployments — wraps the same surface in OpenClaw’s plugin schema.
Git cloneClone the repo to the agent runtime’s skills directoryPower-user path. Useful when you want to fork + customize a skill (override prompts, add WEO-specific wrappers).

Auth flow

API key mode — direct billing. Operator sets an API key env var; HeyGen bills per call. Same unit-economics as any direct-API integration.

MCP / OAuth mode — runs against the operator’s existing HeyGen plan credits. Significant for operators who already pay for a HeyGen subscription and don’t want to manage a separate API-key budget. Same “subscription-as-cost-control” pattern as the Hermes Codex runtime running OpenAI turns against a ChatGPT subscription.

MCP / OAuth setup, as demonstrated (2026-09-29 refresh). Anastasia Martiniuk walked through the connector path live in the HeyGen Spain webinar. She paired HeyGen’s MCP connector with her own downloadable skill folder (short-form “Yap” talking-head skills), not the heygen-com/skills bundle, so this documents the MCP/OAuth mode the README names rather than the bundle itself.

  • No API key. In her words, what you connect “is not an API as such, it’s a connector”: the connection is tied to your HeyGen account when you link it.
  • Setup in the Claude desktop app:
    1. Code tab → ”+” → Connectors → Add custom connector.
    2. Paste HeyGen’s MCP URL (she found it by searching “HeyGen MCP Claude Code”).
    3. Complete the OAuth redirect to HeyGen and authorize access.
    4. Set tool permissions. She chose “always allow” for every tool, including write and delete.
  • Give Claude the IDs before the script. Copy the avatar-style ID and the voice ID from the “three dots” menu on each in HeyGen, then tell Claude which avatar and voice to use before pasting or generating the script.
  • The avatar sticks within a conversation. Once an avatar is in use, Claude keeps it for later requests in the same chat. She rendered two videos with the wrong avatar by not naming the one she wanted.
  • Renders took about 3–5 minutes by her estimate.
  • Cowork is unconfirmed. Asked whether a Cowork pipeline could do this, she was unsure Cowork can reach the connector: “solamente funcionó con Claude Code” (it only worked with Claude Code). That is presenter uncertainty, not a tested limit.
  • Plan credits. Asked whether a basic HeyGen plan is enough, she said it runs on credits consumed by video length and did not give a figure.

How it fits the existing HeyGen + Higgsfield ecosystem in this wiki

SurfaceWhat it isWhere it fits
HeyGen Avatar V (entity)The avatar model itself — video-reference conditioning, 1080p, unlimited duration, 175+ languagesFoundation layer for heygen-avatar skill
Claude Code (Nate Herk)Three-tool pipeline (ElevenLabs + HeyGen + Remotion) hand-orchestrated by Claude CodeOperator-assembled predecessor to the heygen-video skill
heygen-com/skills bundle (this article)Vendor-published 3-skill bundle (avatar / video / translate)Productized agent surface — what Nate Herk’s pipeline was teaching by hand, HeyGen now ships as install-and-go skills
HeyGen HyperframesSeparate repo — HTML composition framework + the /hyperframes router skill bundle (creation workflows + -core/-animation/-creative/-media/-cli/-registry domain skills)Composition layer that pairs with this skills bundle for full-pipeline coverage
skillsSister vendor’s skills bundle (Higgsfield’s four skills)Pattern peer — same “vendor-published agent skills” model in the creative-AI vertical

Routing decision: use this skills bundle when the operator wants HeyGen’s three core production capabilities as agent-callable surfaces inside Claude Code / Cursor / Codex / etc. Use Hyperframes when the composition layer is the bottleneck (captions, transitions, motion graphics, layout). Use both together for the end-to-end pipeline.

Try It

  1. Confirm gh skill is installed. gh skill --help. If not present, install the gh skill extension first (it’s an official GitHub CLI extension).
  2. Install the skill(s) you need. For a typical full-pipeline workflow:
    gh skill install heygen-com/skills heygen-avatar
    gh skill install heygen-com/skills heygen-video
    gh skill install heygen-com/skills heygen-translate
  3. Pick your auth path. If you already pay for HeyGen, configure MCP/OAuth against your plan credits. Otherwise set the HEYGEN_API_KEY env var.
  4. Spin up an avatar. From Claude Code (or whichever runtime), invoke the skill with a photo + a name. The skill returns a persistent identity reference you can reuse.
  5. Generate a first video. Use the persistent avatar reference + a one-line brief. Inspect the script the skill returns; iterate on the brief if the script needs adjustment.
  6. Localize for one extra language. Run the heygen-translate skill on the finished video. Validate the voice-clone + lip-sync quality before scaling to additional languages.
  7. Compose with Hyperframes. Take the rendered video and use hyperframes` to add a caption overlay, lower-third, or end-card. Both repos are MIT and target the same agent runtimes.
  8. Compare cost vs Studio Automation pipeline. Run the same brief through Nate Herk’s three-tool pipeline and through the skills bundle; measure setup-time-to-first-render, per-render credit cost, and operator-control granularity. Different operators will optimize different axes.

Open Questions

  • Skill granularity — three skills cover avatar / video / translate. Is HeyGen planning a fourth (or fifth) skill — for example a heygen-thumbnail (export-frame-as-thumbnail), a heygen-clip (cut-and-export-segment), or a heygen-streaming-avatar (real-time avatar driver)? Worth tracking the releases page on the repo for v3.x and v4 trajectory.
  • MCP/OAuth flow specifics — README confirms the dual-auth option exists but doesn’t document the OAuth scope or token lifecycle. Worth a primary-source pull on HeyGen’s developer docs for the exact contract before relying on it for production deployments. Partly answered 2026-09-29: the user-side setup (custom connector, OAuth redirect, per-tool permissions) is now documented from a practitioner demo in § Auth flow. Scope and token lifecycle are still undocumented.
  • Plugin marketplace fragmentation — the README cites three install paths (gh skill, clawhub, openclaw plugin). Each implies a different marketplace surface tracking the same skills. Worth understanding which one HeyGen actually treats as canonical — pattern hygiene for vendors shipping the same skill via multiple marketplaces.
  • Skills + Hyperframes integration story — both repos are HeyGen’s, both target the same agent runtimes, but the README of heygen-com/skills doesn’t explicitly mention Hyperframes as a composition partner. Open question: is HeyGen positioning these as separate products or as a planned bundle? An eventual heygen-compose skill that wrapped Hyperframes would close that gap.
  • Anthropic-skills-repo cross-listing — does HeyGen plan to mirror these skills into the skills registry so Claude Code users can discover them through Anthropic’s skill marketplace too? Discovery surface matters for adoption.
  • Pricing for the MCP/OAuth path — running against HeyGen plan credits is operationally cleaner than API-key billing, but the per-call credit cost vs API-key pricing isn’t documented in the README. Worth measuring on a small test workflow. The 2026-09-29 webinar confirms only that plan credits are consumed by video length; still no per-call figure.