What an AI draft can and cannot do in NioConnect

Who does the writing

The writing is done by the assistant you connect, not by NioConnect. What NioConnect provides is a door: a set of tools through which an assistant can read your workspace and save posts into it, using the same services the web app uses.

That distinction shapes everything below. An assistant is a participant with a narrow set of abilities, and every post it creates is subject to the rules every other post is subject to. For the list of tools themselves, see connect your AI client over MCP; for the controls around them, see connecting an AI client without losing control.

What an assistant can do

Inside the workspace it was connected to, an assistant can:

• See your connected channels and whether each connection is healthy.
• Look up what a platform accepts: length, media, threads and polls.
• Read posts on the calendar, scheduled and published, including posts waiting for a teammate’s approval.
• See the workflows that copy one channel to another, and look at one in detail.
• Save a post as a draft.
• Schedule a post, when it is given an explicit time that is still in the future.

What it cannot do

The limits matter more than the abilities, so they are worth stating flatly.

• Publish now. There is no publish now option, no argument for it and no second tool for it.
• Create or change a workflow. Workflows can be listed and inspected, not written.
• Read another workspace. Every tool sees the one workspace the connection belongs to.
• Act above its role. It acts with the role of the member who approved it or created its key, checked on every call.
• Skip a check. Its posts pass the same validation, approval, quota and publish admission checks as yours.

Approval, quota and plan limits still apply

If your workspace has approvals on, a post made through an assistant waits for review like any other. If it is over the plan’s allowance, it is refused like any other. The assistant is not a second path with looser rules; it calls the same code the composer calls.

That is also why the Reddit and Discord safeguard holds. Posting into a community defaults to needing approval whatever the workspace setting, and an assistant does not get an exemption. The reasoning is in approval before queueing, explained.

Reviewing a draft before you send it

A draft is cheap to read and cheap to throw away, so read it properly. Open it in the composer and go through the same few things you would check on your own work.

1. Read the preview for each channel, not just the shared text. A draft can be fine on one platform and too long for another, and the checks will say which.
2. Check every link. Open it, and look at how it will unfurl with the link preview check.
3. Look at the images. Make sure each one has a description; the alt text checker is a quick way to see what is missing.
4. Check the time and the time zone, if it was scheduled.
5. Look for the small marker that shows a post was created by an AI client. It is there so a teammate can tell at a glance what a person did not write themselves.

Prompts that work

Plain requests work best, and a good habit is to make the assistant look before it writes. Say what you want, name the channels, and say whether it may schedule.

Check my channels and tell me if any need attention.
What does Mastodon accept in a post, and how long can it be?
Draft a launch post for Bluesky and Mastodon. Leave it as a draft.
What is waiting for approval this week?

Set-up steps for each client are on the AI and MCP pages, with a worked one for Claude Code. A client that takes a key needs one made first: see how to create and revoke an API key.

A session, start to finish

Here is what a careful session looks like, so the pieces above sit in order. You ask the assistant for a launch post for two channels and tell it to leave the result as a draft.

1. The assistant lists your channels to find the two it needs, and notes if either connection is unhealthy.
2. It looks up what each platform accepts, so it writes inside the limits rather than discovering them by being refused.
3. It saves one post to both channels with no time. If the text were too long for either, the tool would refuse and name the limit, and nothing would be saved.
4. You open the draft, read each channel’s preview, fix what needs fixing and schedule it yourself, or send it for review if your workspace requires that.

Two identical requests make two posts, because each call creates one. If something looks doubled, check the calendar before asking again.

Who is responsible

You are. An assistant draft that reaches your audience reached it because someone with the right role let it through, whether by approving it, scheduling it or switching approvals off. The audit log records who approved what, so that question always has an answer.

If you would rather an assistant never have the option, the safest setup is the narrowest one: connect it with a member whose role cannot publish without review, and turn approvals on for that role.