YouTube Automation Agent

The open-source AI agent that runs a YouTube channel end to end, with you approving every video before it goes out.

It researches topics, writes scripts, generates narration and visuals, assembles real MP4s, writes titles, descriptions and tags, schedules approved videos, and learns from analytics and from what your audience says. It runs on your own machine with your own API keys.

Latest release: v2.11.0 (October 13, 2026). See the changelog for release notes and what has landed on master since.

Why it's different

• You approve before anything is published. Finished videos wait for fact review, media-rights confirmation, and your approval. Turning on autonomy doesn't skip those checks.
• Real output only. Simulated or placeholder media is clearly marked and can never be approved or uploaded.
• Self-hosted. Your keys, media, and channel data stay on your machine.
• Learns from evidence, with your sign-off. Lessons from analytics, retention curves, and comments shape future videos only after you approve them.
• Your choice of providers. Eight text providers or OpenRouter, five AI video providers, or free local slideshow rendering.

Quick start

You need:

• Node.js 18.17 or newer
• A Google account and YouTube Data API credentials (setup steps)
• An API key for at least one text provider. Gemini has free tiers for text and narration.

git clone https://github.com/darkzOGx/youtube-automation-agent.git
cd youtube-automation-agent
npm install
npx playwright install chromium   # renders slideshow videos; on Linux add --with-deps
npm run walkthrough
npm start

Then open http://localhost:3456.

1. npm run walkthrough explains each provider choice, tests your keys, and connects your YouTube channel. For a shorter flow, use npm run setup. Every setting is documented in .env.example.
2. Check readiness. In the dashboard, open Production readiness and choose Run verified check. It makes small live text and narration requests, checks access to your channel, and builds and decodes a test MP4, without uploading anything.
3. Make videos. Open Autonomous operator, describe the channel you want, and choose Activate & run now, or create a single video from the dashboard. Finished videos wait in Pipeline for your review and approval.

FFmpeg is installed with npm install through ffmpeg-static. Python is only needed to test an external DarkzSEO checkout.

How it works

Stage	What the agent does	What you control
Research	Finds topics and builds a content strategy	Niche, audience, blocked topics
Script	Writes the hook, narrative, CTA, and metadata	Voice, format, length, brand direction
Production	Generates narration and visuals, then assembles a real MP4	Provider choice and media fallbacks
Review	Runs quality checks and opens the video for review	Facts, media rights, edits, approval
Publish	Schedules and uploads approved content	Privacy, timing, final decision
Learn	Captures 24-hour and 7-day evidence, measures your chosen outcome, then proposes the next move	Choose the KPI and approve or reject each lesson

graph TD
    O[Autonomous Channel Operator] --> A[Research and Editorial Plan]
    A --> B[Content Strategy Agent]
    B --> C[Script Writer Agent]
    C --> D[Thumbnail Designer Agent]
    C --> E[SEO Optimizer Agent]
    D --> F[Production Management Agent]
    E --> F
    F --> Z[DarkzSEO Discoverability Preflight]
    Z --> G[Review and Approval Gates]
    G --> H[Publishing & Scheduling Agent]
    H --> I[Analytics & Optimization Agent]
    I -->|feedback loop| A

Features

Each item links to its walkthrough in the feature guide.

Plan

• Autonomous Channel Operator: give it an objective, audience, content pillars, cadence, and limits. It researches trends and competitors, builds an evidence-labeled plan, and sends each video through production.
• Outcome & ROI Studio: set a target (views, watch hours, net subscribers, engagement, or revenue) and a budget, and track progress. Missing data shows as unavailable, never as zero.

Produce

• Production readiness: a verified check of every provider, your channel, and video assembly before anything runs on its own. Paid image and video probes are separate opt-ins.
• Resumable generation: every stage is checkpointed, so a crash or timeout resumes where it stopped instead of starting over.
• AI video generation: local slideshow by default; Seedance, MiniMax H3, Gemini Omni, Kling, or Wan for real clips, with a cap on paid seconds.
• Shorts Repurposing Studio: turn an approved video into three 9:16 Short drafts with burned-in captions, reusing existing media so no new generation credits are spent.

Review

• Content review: quality checks, metadata editing, approval, and scheduling for each video.
• Scene Repair Studio: edit, reorder, lock, replace, or regenerate a single scene, then rebuild the video.
• Evidence desk: link each factual claim in the script to a verified source before approval.
• Discoverability preflight: advisory search and AI-answer checks from a bundled DarkzSEO audit. No Python needed.

Learn

• Learning loop: compares each video's 24-hour and 7-day results with your channel's own history and proposes changes for you to approve.
• Controlled growth experiments: test approved title and thumbnail variants on a live video. A winner needs a 95% evidence threshold, and the original is restored before you decide.
• Scene-aware retention: see which scene lost viewers.

Engage

• Audience Engagement Studio: comment sync, themes and questions, spam quarantine, reply drafts that post only after you approve them, and video ideas mined from repeated requests.

AI providers

Text

Pick one in the walkthrough, or set its key in .env. The SDK base URL is configured for you.

Provider	Models	Key	Env var
Google Gemini	Gemini 3.8 Flash, Gemini 4 Argon (paid only), 3.5 Flash-Lite	AI Studio	GEMINI_API_KEY
OpenAI	GPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna	platform.openai.com	OPENAI_API_KEY
OpenRouter	400+ models through one key; curated defaults are checked against its live catalog	openrouter.ai	OPENROUTER_API_KEY
Kimi (Moonshot AI)	Kimi K3, K2.7 Code, K2.6	platform.kimi.ai	MOONSHOT_API_KEY
MiMo (Xiaomi)	MiMo V2.6 Pro, V2.6 Flash	mimo.mi.com	MIMO_API_KEY
GLM (Zhipu AI)	GLM-5.3, 5.3 Flash, 5.2	z.ai	GLM_API_KEY
DeepSeek	DeepSeek V4.1 Flash, V4 Pro	platform.deepseek.com	DEEPSEEK_API_KEY
MiniMax	MiniMax M3.1 Flash, M3	platform.minimax.io	MINIMAX_TEXT_API_KEY

Anthropic Claude is available through OpenRouter (anthropic/claude-fable-5.1). OpenRouter, Kimi, MiMo, GLM, DeepSeek, and MiniMax cover scripts only. Add a Gemini or OpenAI key for images and narration, or ElevenLabs for narration.

Narration and images

• Narration uses the first configured provider in this order: ElevenLabs (Eleven v3, needs ELEVENLABS_API_KEY and ELEVENLABS_VOICE_ID), OpenAI (gpt-4o-mini-tts), then Gemini (gemini-3.8-flash-tts). A video without working narration can't be approved.
• Images use OpenAI (GPT Image 2) and Gemini (gemini-3.1-flash-image, which needs billing enabled), each falling back to the other. With neither, slides use gradient backgrounds and thumbnails come from the built-in designer.

Video

Local slideshow rendering is the default, so you never pay for video unless you choose a provider in Channel setup and set a paid-seconds cap. If a provider fails, the video falls back to the local slideshow.

Provider	Default model	Best fit	Clip limits	Env var
ByteDance	bytedance/seedance-2.5 through Replicate	Cinematic long scenes and large reference sets	4–30 seconds	REPLICATE_API_TOKEN
MiniMax	MiniMax-H3	Multimodal references, native stereo audio, optional 2K	4–15 seconds	MINIMAX_API_KEY
Google	gemini-omni-1.1-flash	Fast generation and conversational editing	3–10 seconds	GEMINI_API_KEY
Kuaishou	kling-v3-omni	Storyboards and character/voice consistency	3–15 seconds	KLING_ACCESS_KEY, KLING_SECRET_KEY
Alibaba	Wan 2.7 task-specific models	Efficient generation, reference video, and continuation	2–15 seconds	DASHSCOPE_API_KEY

Long videos use hybrid assembly: generated clips for the hook and key sections, with the rest of the timeline filled locally. See AI video generation.

Configuration

YouTube (required, free)

1. Create a project in Google Cloud Console.
2. Enable YouTube Data API v3, and YouTube Analytics API for retention and outcome data.
3. Create an OAuth 2.0 client of type Desktop app (not Web application) and save the JSON as config/credentials.json.
4. Under Google Auth Platform → Audience, add the Google account that owns your channel as a test user. Otherwise sign-in fails with Error 403: access_denied. While the app is in Testing mode, Google expires the sign-in after 7 days.

Environment variables

npm run walkthrough writes your configuration for you. To edit it by hand, copy .env.example to .env; it documents every setting. The essentials:

# One text provider (see the table above)
GEMINI_API_KEY=...

# Optional: narration, images, and AI video
# ELEVENLABS_API_KEY=...
# ELEVENLABS_VOICE_ID=...
# OPENAI_API_KEY=sk-...
# VIDEO_PROVIDER=slideshow   # auto, seedance, minimax_h3, google_omni, kling, wan
# VIDEO_MAX_GENERATED_SECONDS=60

# App
PORT=3456
# HOST=127.0.0.1             # default: this machine only
# API_KEY=some-long-random-string
CHANNEL_NAME=Your Channel Name
DEFAULT_PRIVACY_STATUS=private

The dashboard listens on 127.0.0.1 by default. If you set HOST=0.0.0.0 to reach it from other devices, also set API_KEY, which every request that changes something must then send in the x-api-key header.

Automation

The scheduler starts with npm start and uses the server's local time.

When	What runs
Daily at 06:00	Content generation, following the active channel strategy's cadence
Every 15 minutes	Publishing queue
Daily at 09:00	Analytics collection
Daily at 22:00	Optimization
Every 4 hours	Comment sync and growth-experiment refresh
Sundays at 08:00	Weekly strategy review

Details are in the feature guide.

HTTP API

The dashboard server also exposes an HTTP API:

curl http://localhost:3456/health

curl -X POST http://localhost:3456/generate \
  -H "Content-Type: application/json" \
  -H "x-api-key: $API_KEY" \
  -d '{"topic": "Top 10 Life Hacks", "style": "list"}'

curl http://localhost:3456/api/jobs/:jobId

See the API reference for every route.

Troubleshooting

Problem	Fix
Missing credentials for: an AI provider	Configure any one text provider with npm run walkthrough or npm run credentials:setup. OpenAI isn't required.
Google sign-in shows Error 403: access_denied	Add your Google account as a test user under Google Auth Platform → Audience.
No video is produced, or it's marked simulated	Read the ✗ lines in the startup capability check. Make sure you ran npx playwright install chromium and that FFmpeg is available (npm install fetches it, or set FFMPEG_PATH).
Images or thumbnails are missing	Image generation needs OpenAI, or Gemini with billing enabled. The content review panel shows the provider's exact error.
Narration failed (for example, ElevenLabs HTTP 400)	Check the key, its credits, and that ELEVENLABS_VOICE_ID is a voice in your account, then use Regenerate narration only.
Error loading sharp module on Android/Termux	Update to the latest master, then run rm -rf node_modules && npm install --cpu=wasm32.
The publishing queue runs but nothing publishes	Approved videos publish at their scheduled time (default: the next day at 2 PM).
YouTube API quota exceeded	Check quotas in Google Cloud Console and reduce posting frequency.
Content generation failed	Check your keys and credits, the job's error in the dashboard, and logs/.

For debug logging, run LOG_LEVEL=debug NODE_ENV=development npm start. Questions and setup help go in Discussions.

Project structure

youtube-automation-agent/
├── agents/          # one file per pipeline agent (strategy, script, thumbnail, SEO, production, publishing, analytics)
├── dashboard/       # web dashboard served by index.js
├── database/        # SQLite schema and access layer
├── schedules/       # cron-based automation
├── utils/           # AI text, media, and video providers; operator, readiness, engagement, and other services
├── config/          # credential files (examples are committed)
├── data/            # generated content and the local database
├── docs/            # feature guide, API reference, design notes
├── examples/        # verified real output from the community
├── index.js         # Express server and agent startup
├── walkthrough.js   # guided setup (npm run walkthrough)
├── setup.js         # classic setup (npm run setup)
└── test.js          # system test suite (npm test)

Extending

Add a text provider

Any provider with an OpenAI-compatible chat completions API can be added as a preset in utils/ai-text-service.js:

deepseek: {
  name: 'DeepSeek',
  baseURL: 'https://api.deepseek.com',
  defaultModel: 'deepseek-flash',
  models: ['deepseek-flash', 'deepseek-v4-pro'],
  envKey: 'DEEPSEEK_API_KEY',
},

Then add a matching entry to AI_PROVIDER_GUIDE in walkthrough.js and a setup option in utils/credential-manager.js, and run npm test. The walkthrough test fails if any provider is missing a complete guide.

Add a video provider

Video providers live in utils/video-providers.js. Each one extends VideoProvider, implements isAvailable, createTask, getTask, and downloadResult, and is registered in VideoProviderRegistry.

Contributing

See CONTRIBUTING.md for ground rules (short version: one focused concern per PR, no lockfile churn, lint and tests must pass). For questions and setup help, use Discussions; Issues are for bugs.

git clone <your-fork>
cd youtube-automation-agent
npm install
npm run lint   # CI runs this on every PR
npm test

Made something real with it? Share verified output in examples.

Built by

@darkzOGx, a solo builder shipping AI automation and developer tools. Find me on X and laderalabs.io.

If YouTube Automation Agent saves you time, a star helps it reach more developers.

More from darkzOGx:

• darkzloop: terminal agent runner that turns any LLM into a disciplined software engineer (FSM control, model-agnostic, BYO auth)
• darkzBOX: open-source Instantly.ai clone with smart automated email replies
• open-sales-researcher: autonomous B2B company research. Works with Claude Code, Cursor, Copilot.
• darkzseo: SEO tooling
• ConstructionBids.ai: AI scans every federal, state & local public works bid and matches you to contracts you'll win.

License

MIT. See LICENSE.

Acknowledgments

• Google: Gemini 3.8 Flash, Gemini 4 Argon, Gemini 3.1 Flash Image, Gemini 3.8 Flash TTS, Gemini Omni Flash, and the YouTube Data and Analytics APIs
• OpenAI: GPT-6.1 Sol, GPT Image 2, gpt-4o-mini-tts
• OpenRouter: unified multi-model API
• Moonshot AI: Kimi K3
• Xiaomi: MiMo V2.6 Pro
• Zhipu AI: GLM-5.3
• DeepSeek: DeepSeek V4.1 Flash
• MiniMax: MiniMax M3 text and H3 video
• ElevenLabs: Eleven v3 TTS
• Replicate: Seedance 2.5 and Wan 2.7 hosting
• Kuaishou Kling: Kling 3.0 Omni
• Alibaba Cloud Model Studio: Wan 2.7

---

This tool is for legitimate content creation. Comply with YouTube's Terms of Service and Community Guidelines.

---

TalentDelta provenance record (added at packaging)

• Packaged for: TalentDelta catalogue (draft; gated by review)
• Upstream: darkzOGx/youtube-automation-agent
• Pinned commit: 24f9cb521fcbebec70b7a193e0f6d972aeec88a6 (2026-10-20)
• Licence: MIT — MIT text at the repository root, shipped byte-identical to the pin, copyright '2025 YouTube Automation Agent Contributors'. Permissive: use, modification and redistribution allowed, including commercially, provided the copyright notice and permission notice travel with substantial portions. No NOTICE file upstream and no per-file headers are added; attribution is recorded in the pack README's provenance section and in this upload record.
• Upstream author credit: darkzOGx
• Runtime: Node.js 18.17+ (npm install, npx playwright install chromium for slideshow rendering, npm run walkthrough, npm start); dashboard on http://localhost:3456; YouTube Data API credentials plus at least one text-provider key; FFmpeg ships via ffmpeg-static
• Changes made at packaging: this section was appended to README.md. No upstream path is excluded. Apart from this provenance section appended to README.md, no file was added, removed, renamed or edited; every upstream file is byte-identical to the pinned commit.
• Safety read of the packaged archive (packaging time, 2026-10-24): Static safety read of the shipped archive (71 files): 15 pattern hits, none a live credential. The 14 api-key patterns are environment-variable reads or credentials-object fields in code, plus the README's placeholder line; the 1 address hit is an npm deprecation notice inside package-lock.json. Two dummy key strings in test.js are 15 characters long, shorter than any real provider format. No credential, token or private key ships; .env.example holds empty values.

Packaging notes (TalentDelta)

• Seven agents are real modules with their own prompts and tool use, not a prompt library: content strategy, script writing, thumbnail design, SEO optimisation, production management, publishing/scheduling and analytics optimisation, all constructed in index.js.
• Publishing is gated by design: finished videos wait for fact review, media-rights confirmation and explicit approval, and simulated or placeholder media is marked so it cannot be approved or uploaded.
• No credentials ship. .env.example holds empty placeholders; the YouTube OAuth client secret, provider keys and channel tokens are supplied by the operator and stored locally through the credential manager.
• Self-hosted: API keys, generated media and channel data stay on the operator's machine. The anonymous-telemetry module ships disabled: it sends nothing unless ANONYMOUS_TELEMETRY_ENABLED is set to true and an https endpoint is configured, and it never sends content.
• Upstream states that AI-generated narration and visuals must be rights-checked before publishing; that guidance is preserved verbatim in the shipped README.