Giving Claude or ChatGPT a connector to your social accounts saves real time: "write the launch post and schedule it for Friday 9am" becomes one sentence. The worry is just as real. An agent that can post can also post a line you never read, under your name, to everyone who follows you.
You do not have to choose between the two. There are three layers of approval you can stack, and each one stops a different mistake. Disclosure up front: we build Posteora, a posting connector for AI assistants, and the examples use it. The first layer works with any connector.
Layer 1: make the assistant ask before it calls a posting tool
Every MCP tool tells the assistant whether it only reads or also changes something. Posteora marks its publish and schedule tools as write actions, so your assistant's own permission settings apply to them.
- Claude Code: add the posting tools to the
asklist in.claude/settings.json, and Claude Code stops for your yes every time:
json
{
"permissions": {
"ask": ["mcp__posteora__post_publish", "mcp__posteora__post_schedule"]
}
}
- Claude (claude.ai and Desktop): in the connector's settings, set the posting tools to need approval instead of always allowed.
- ChatGPT (developer mode): ChatGPT asks you to confirm write actions before it runs them. Read the confirmation; do not click through it.
This layer stops the agent from acting on its own. It does not show you a clean preview of the post, which is the next layer.
Layer 2: preview first, with a receipt
Posteora's instructions tell the assistant to call post.preview before anything is queued. The preview returns the text as each network will take it, the character fit, the time it will go out, link warnings for X, and what it draws from your plan. The assistant reads that back to you and waits.
A preview is free and changes nothing, so you can ask for three versions and pick one.
This layer depends on the assistant following its instructions. Layer 1 and Layer 3 do not.
Layer 3: drafts only, approved in the Studio
For the strongest version, ask the assistant to save drafts instead of scheduling. With draft set, Posteora saves the post to the Drafts tray in the Studio, and nothing is queued until you approve it there yourself.
A good standing instruction for your assistant:
Write and preview posts, but always save them as drafts. Never publish or schedule directly.
Then once a day, open Drafts, edit what needs editing, and approve the rest. The assistant does the writing; the publish button stays yours.
Afterwards: you can still pull it back
Everything an assistant schedules appears in your Studio calendar like any other post, so you can edit or cancel it before it goes out. Every tool call is kept in an audit log, and you can sign any assistant out from the AI Assistants page in your dashboard.
Which layers to use
- Trying it out: all three. Ask on every tool call, read every preview, save as drafts.
- Once you trust the voice: keep Layer 1 on publish and schedule, drop drafts for routine posts.
- For a team or a client account: drafts only, always.
Posteora is free to start: 10 posts a month across X, LinkedIn, Bluesky, Mastodon, TikTok, Pinterest and YouTube, with the assistant connection included. Setup for each assistant is on the MCP docs.
Related: how to post to LinkedIn from Claude and how to post to X from ChatGPT.