An AI agent that actually asks before it tweets

What it does

This LangGraph agent ingests a URL, scrapes the content, and drafts posts for Twitter and LinkedIn. Before publishing, it pauses for human review so you can edit or reject the output. It also handles the auth lifecycle for social platforms, either through Arcade or by running local OAuth servers for Twitter and LinkedIn.

The interesting bit

The real challenge is not the copywriting — it is safely automating access to corporate social accounts without letting an LLM post unsupervised. The workflow treats authentication, media uploads, and platform scheduling as core concerns, making human approval a deliberate architectural chokepoint rather than an afterthought.

Key highlights

• Built on LangGraph with a workflow that separates content scraping, generation, human review, and scheduling.
• Offers a minimal quickstart for generic URLs, plus a full setup that adds GitHub, Twitter, YouTube, Slack, and image upload support.
• Human-in-the-loop approval via Agent Inbox before posts are scheduled to go live.
• Social authentication abstracted through Arcade, with fallback to manual Twitter and LinkedIn developer app OAuth.
• Customizable prompts and post styles for adjusting tone across platforms.

Caveats

• Full features require a wide array of API keys and services: Anthropic, FireCrawl, Arcade, Google Vertex AI, Supabase, and separate Twitter or LinkedIn developer accounts for media uploads.
• The quickstart explicitly strips out parsing for GitHub, Twitter, and YouTube, plus Slack integration and image handling — it is a text-only sandbox.
• The README is dominated by environment variable and OAuth configuration, so most of the upfront work is auth plumbing, not content tuning.

Verdict

Useful if you are prototyping agentic publishing pipelines and need a working pattern for human-in-the-loop auth and approval. Skip it if you want a lightweight, drop-in social media cross-poster — this is an integration-heavy framework demo, not a polished product.