Explanation
How assistant connections work
The shared address, which Claude's connectors directory uses, and your workspace's own address, what each lets an assistant do, what SemPost receives and keeps, what protects your workspace, and assistants marked Unverified.
An AI assistant such as Claude works in SemPost through a connection. Each connection works in one workspace only, with the role of the owner who made it and the permissions that owner ticked on SemPost's approval page. The assistant never gets your password, and it can never do more than your role in the workspace allows.
You make a connection with one of two addresses. Both reach the same workspace and the same data; what differs is how you choose the workspace and what the assistant may ask for.
The two addresses
| The shared address | Your workspace's own address | |
|---|---|---|
| The address | https://mcp.sempost.ai/mcp, the same for everyone | https://mcp.sempost.ai/mcp/ followed by the workspace's own key, in Settings › Connected assistants |
| Where you add it | In Claude, from SemPost's listing in Claude's connectors directory; in any assistant, as a custom connector | In your assistant, as a custom connector; for Claude, Add to Claude in Settings › Connected assistants fills it in |
| The workspace | You choose it on SemPost's approval page, under Which workspace? | Always that address's workspace |
| Reading, writing posts, settings, media, approving posts, starting RSS feeds, deleting on your confirmation | Yes, with the permissions you tick | Yes, with the permissions you tick |
| Approving X posts, which spend credits as they go out | No: you approve them in SemPost | Yes, each on your confirmation; the page shows each X post's credits |
| Switching automation rules on, resuming automation, taking an account back into automation, removing days without automation posts | No: you do these on the Automation page | Yes, each on your confirmation |
| Video Studio projects: setting them up, editing their shots, cuts and recipes, free drafts, deleting them | No: it reads your projects and can call off a draft or final that's still waiting | Yes |
| AI work that spends credits: scenes, new images, posts, Studio videos, Video Studio finals | Never | Yes, where it's open to your workspace and you tick the box that starts with Ask for work that spends credits, each at its price on your confirmation |
Switching a rule on starts it preparing paid posts, and a Video Studio cut is what a paid final is made from. That's why both stay with the workspace's own address: a connection made with the shared address never spends credits, not even through a rule or an X post, and never makes AI pictures, videos or sound.
Which one to use
- For your own work in SemPost, use the workspace's own address. It can do everything an assistant may do in SemPost, AI work too when you allow it, and you don't pick the workspace as you connect.
- For an assistant that must never spend credits or make AI media, use the shared address. Everything else still works: it reads your workspace, writes posts that wait for your approval, changes the settings you allow and asks for your confirmation before approving or deleting anything.
- In Claude, use the connectors directory for the shared address. SemPost's connector there, https://claude.ai/directory/sempost, uses the shared address and works as it does; Claude shows it with a Community label (The Community label). For the workspace's own address, press Add to Claude in Settings › Connected assistants.
Using both
Claude on the web, on the desktop and on your phone counts as one assistant for SemPost. If you connect it to the same workspace with both addresses, the newer connection replaces the older one, and Settings › Connected assistants lists only the new one. Claude doesn't take the same address twice either. So for one workspace, keep the connector with the address whose work you need, and remove the other one from Claude.
An assistant that runs on your own computer, such as Claude Code, keeps one connection per computer, whichever address it uses: connecting it doesn't end your connection in Claude on the web, and connecting Claude on the web doesn't end it.
What SemPost receives and keeps
- Your conversation stays with your assistant. SemPost gets only the requests the assistant sends, such as "list the posts waiting for approval" or "write this draft", and answers them.
- What the assistant reads goes to its provider, for example Anthropic for Claude, which handles it under its own terms with you.
- The workspace's audit log records every connection made or ended and every change an assistant makes, with the assistant's name (The audit log).
- Requests that wait for your confirmation are kept with their answers for 30 days after they were made, whatever their outcome.
- The day after an assistant changes something, the workspace's owners get one email about that day.
An assistant doesn't see the team's names, email addresses or roles, invitations, the audit log, payment details, keys or passwords, or the media files themselves.
What protects your workspace
- Owners decide. Assistant connections are off until an owner turns on Allow assistant connections in Settings › Workspace, and only owners connect assistants. Turning the setting off ends every connection at once.
- You choose the permissions. You approve each connection signed in to SemPost with your password, and tick what the assistant may do. Reading comes with every connection; the rest is your choice.
- Posts wait for a person. Every post an assistant writes or changes waits in the Approval queue, marked Assistant, even when your approval rule is None. SemPost checks once more before a post goes out.
- The confirmation page. Approving and planning posts, starting an RSS feed, anything that can't be undone and work that spends credits run only after you confirm on SemPost's own page, signed in to your account. The page shows the posts, rules and files as they are in SemPost, not as the assistant described them. Words the assistant wrote that would go into AI work, such as a voiceover, are quoted in a box of their own, What the assistant wrote in the request, and a rule you're asked to switch on shows What the rule's posts will carry: SemPost uses these words as they are, so read them before you confirm. If you didn't ask for something, decline it: an assistant may follow an instruction it read in a post, a product, a file's name or a feed.
- Limits. Assistants write at most 100 posts in a workspace in 24 hours, a connection has at most 5 requests waiting at once, and SemPost limits how often each connection can call it.
- Paid work. Paid work goes through the same content check as the panel's, within the assistants' daily credit allowance and your budget (AI work through an assistant).
- You can end it at any time. Press End connection in Settings › Connected assistants. Removing SemPost from your assistant doesn't end the connection on SemPost.
Assistants marked Unverified
When an assistant names a document on its own domain that confirms who it is, SemPost's approval page shows its name with that domain, for example "Claude (claude.ai) wants to connect to SemPost", marked Verified domain. Claude and Claude Code connect this way.
Some assistants register themselves instead. The name they give is their own claim, so SemPost shows it marked Unverified, on the approval page and in Settings › Connected assistants. Such an assistant isn't necessarily unsafe, but SemPost can't check its name. Approve only a connection you started yourself from an assistant you know, and check where the approval goes: the page says The approval goes to … with the assistant's address. If you didn't start it, press Deny.