HeyGen CLI

HeyGen CLI

Make AI videos from the command line. Drive HeyGen with code, not clicks.

Full reference and examples: developers.heygen.com/cli.

Built for

• Coding agents — Claude Code, Codex, and others
• CI/CD pipelines — weekly recap videos, release-note vlogs
• Bulk operations — translate 100 videos in one shell loop
• Custom integrations — wrap it in your own tool

Agent skills

heygen-com/skills — one-line install for Claude Code, Codex, and other agents. Create your own avatar and generate a video in a single conversation.

Agent-first by design

• JSON on stdout, structured errors on stderr, stable exit codes.
• Self-describing. --request-schema and --response-schema return JSON Schema without auth or API calls.
• Non-interactive by default. Set HEYGEN_API_KEY and nothing reads a TTY.
• Tell us how it went. After a flow works (or when you hit a bug), run heygen feedback --rating <1-5> --comment "...". It sends an anonymous rating + note (no API key needed); honors the analytics opt-out. Agents should use this for bugs rather than opening GitHub issues automatically (see Reporting bugs).

Install

curl -fsSL https://static.heygen.ai/cli/install.sh | bash

Single static binary, no runtime required. Installs to ~/.local/bin.

Supported platforms: macOS, Linux, and Windows (via WSL).

You only need a HeyGen API key — see Authenticate below.

Updates

heygen update            # install the latest version
heygen update --check    # report whether one is available, without installing

Shell completion

Tab-completion for commands, subcommands, and flags is available for bash, zsh, fish, and PowerShell:

source <(heygen completion zsh)     # current shell (bash/zsh/fish)
heygen completion zsh > "${fpath[1]}/_heygen"   # persist (zsh; adjust path per shell)

Run heygen completion --help for per-shell install instructions.

Authenticate

Choose one of the options below. The first three are agent- and CI-friendly; the remaining options are for humans.

1. Environment variable — agents, CI; ephemeral, no file on disk:

export HEYGEN_API_KEY=your-key-here

2. Pipe to auth login — agents; persists API key to ~/.heygen/credentials:

echo "$KEY" | heygen auth login

3. Explicit --api-key — humans paste from a prompt:

heygen auth login --api-key

4. Browser OAuth — humans, Pro / Max subscription users; persists OAuth tokens to ~/.heygen/credentials:

heygen auth login --oauth

5. Device OAuth — humans in an attended SSH/headless terminal; open the displayed URL on any browser and enter the one-time code:

heygen auth login --device

Device OAuth is refused in CI and HEYGEN_NONINTERACTIVE mode. Unattended agents must use an API key.

6. Interactive picker — humans, no flag: a TTY prompt lets you choose between API key (uses API credits) and OAuth (uses subscription credits):

heygen auth login

Verify any of the above with heygen auth status. When an API key is active, the status includes its name, permission mode and scopes, and expiration. Get an API key at app.heygen.com/settings/api.

Single-credential file. The credentials file holds at most one of api_key / OAuth tokens at any time. Running heygen auth login (any method) clears the other on success — re-login overwrites, it does not merge. heygen auth status will tell you which one is active.

HEYGEN_API_KEY in the environment always wins over either file credential.

Quick start

1. Create a finished video from a prompt (returns JSON including video_id):

heygen video-agent create --prompt "30-second product demo" --wait

2. Get its metadata and share link:

heygen video get <video-id>

Returns JSON with video_url (raw mp4), video_page_url (shareable UI link), thumbnail_url, and duration. Pipe to jq to extract what you need:

heygen video get <video-id> | jq -r '.data.video_page_url'
# → https://app.heygen.com/videos/...

3. Download the mp4:

heygen video download <video-id>

Add --human to any command for a readable layout. Set HEYGEN_OUTPUT=human to make it the default.

In --human mode, list responses render as a table and single-object responses render as an indented, humanized layout (similar to kubectl describe). Keys are humanized per segment (auto_reload → Auto Reload), with common acronyms uppercased (video_url → Video URL); nested objects indent under a Label: header, scalar siblings align locally within each block, arrays of scalars join inline (a, b, c), and arrays of objects render as YAML-style - sequence items. Empty objects, empty arrays, and nulls show as (none). Long scalar values (e.g. pre-signed download URLs) are printed in full and may wrap in a narrow terminal; pipe the default JSON to jq if you need to extract one cleanly. For example:

Status:  completed
Wallet:
  Auto Reload:
    Enabled:  false
  Currency:           usd
  Remaining Balance:  476.78
Tags:  alpha, beta

--human is a readable layout for terminals and may change between releases; scripts and agents should consume the default JSON output, which is stable. JSON output (the default) is never altered.

Commands

Mirrors the HeyGen v3 API. Pattern: heygen <noun> <verb>.

Group	What it does
video-agent	Create videos from text prompts using AI
video	Create, list, get, delete, download videos; batch-create and bulk status
template	List templates and generate videos from them
avatar	List and manage avatars and looks
voice	List voices, design voices, generate speech
audio	Search the background-music catalog
model	Train professional voice clones and synthesize speech from them
video-translate	Translate videos into other languages
lipsync	Dub or replace audio on existing videos
webhook	Manage webhook endpoints and events
asset	Upload files for use in video creation
user	Account info and billing

Every command supports --help.

How it behaves

Aspect	Behavior
stdout	Always JSON. Even video download — binary writes to disk; stdout emits {"asset", "message", "path"} so you can chain on .path.
stderr	Structured envelope on error: {"error": {"code", "message", "hint", "param", "doc_url", "request_id"}}. code/message are always present; hint/param/doc_url/request_id appear when applicable (param/doc_url are surfaced from the API for validation and documented errors). Stable code values for programmatic branching. A code prefixed cli_ is originated by the CLI itself (client/transport/local conditions, e.g. cli_download_url_expired). A bare code is either an API code (or a CLI mirror of one) or one of a small frozen set of legacy CLI codes that predate the prefix. The cli_ prefix is reserved for the CLI, so a new CLI code can never collide with an API code. Non-fatal diagnostics share stderr in the same envelope style and never affect the exit code: {"warning": {"message"}} when an invocation was degraded or used something deprecated, and {"notice": {"code", "message"}} for something you may want to know rather than act on. One compact envelope per line, so a single run can emit several. In --human mode all three render as styled Error:/Warning:/Note: lines instead. This covers diagnostics the CLI emits through its formatter; interactive commands (auth login, destructive confirmations) also write human prose to stderr by design.
Exit codes	0 ok · 1 API or network · 2 usage · 3 auth / not permitted · 4 timeout — a request exceeded its per-operation timeout, or the --wait poll window elapsed (stdout contains partial resource for resume)
Request bodies	Flags for simple inputs; -d for nested JSON (inline, file path, or - for stdin). Flags override matching fields.
Idempotency	Endpoints that accept Idempotency-Key expose --idempotency-key; --help on the command is the authority. The CLI never generates one, so a retry is deduplicated only if you pass the same key you sent originally. Replay is limited to 24 hours, and a key reused after that window creates a second resource, so reconcile a stale operation with its get/list instead of retrying it. Keys are 1-255 characters from [A-Za-z0-9_:.-], and a UUID is a safe default. Scope is per-endpoint and per-resource.
Async jobs	--wait blocks with exponential backoff; --timeout sets max (default 20m). Only some create commands support them — --help on the command is the authority. Poll anything else with whichever get reads the id the create returned. 429 and selected transient 5xx (500/502/503/504) retry automatically on retry-eligible requests.

Example error envelope:

{"error": {"code": "not_found", "message": "Video not found", "hint": "Check ID with: heygen video list", "doc_url": "https://developers.heygen.com/docs/error-codes#not-found"}}

Configuration

File	Purpose
~/.heygen/credentials	API key or OAuth tokens (one at a time — see Authenticate)
~/.heygen/config.toml	Output format and other non-secret settings

HEYGEN_API_KEY and HEYGEN_OUTPUT env vars override the respective files.

Update notice. A release build checks for a newer stable release at most once a day, in the background, without delaying the command. When one exists it prints {"notice": {"code": "cli_update_available", ...}} on stderr (a Note: line with --human), at most once a day, and never changes the exit code. It is off in CI (CI, GITHUB_ACTIONS, HEYGEN_NONINTERACTIVE) and skipped for help, completion, update, --version, and the schema flags. Turn it off with HEYGEN_NO_UPDATE_CHECK=1 or heygen config set update_check false.

heygen config list       # show all settings with sources

Reporting bugs

• Quick signal (recommended for agents): heygen feedback --rating <1-5> --comment "...". Goes to private, anonymous analytics. Safe to run unattended. Usage telemetry is anonymous until you sign in: running heygen auth login links it to your account email or username. A one-time notice prints the first time analytics runs; opt out anytime with HEYGEN_NO_ANALYTICS=1 (or heygen config set analytics false).
• Tracked bug a maintainer should see: open a GitHub issue. This repo is public, so review what you paste first and omit anything sensitive (API keys, prompts, internal URLs, personal data).

Agents: do not open GitHub issues automatically. Use heygen feedback for bug signal; if something seems worth tracking, surface it to the user and let a human file the issue after reviewing it for sensitive content.

Contributing

See CONTRIBUTING.md. Maintainers using a coding agent such as Claude Code or Codex can run the release, E2E, and changelog workflows as skills; see .agents/skills/README.md.

License

Apache License 2.0

Links

• CLI Documentation
• API Documentation
• Releases