AI Community Manager
What it actually does.
The Community Manager wakes up every few minutes, scans tenant state through six generators, and writes drafts into a queue at /admin/community-manager. The drafts wait there until you approve, edit, or reject. There are six kinds:
1. Welcome drafts.
When a new member signs up, the welcome generator reads their stated interests, their first forum subscriptions, and the most-active forums on your tenant, and writes a personalized greeting. It points them at the thread most likely to get them their first reply. You approve. It posts as your community manager account, or DMs them, depending on your setting.
2. Weekly summary drafts.
Every Sunday morning, the weekly summary generator pulls the top threads of the past 7 days, the new-member count, milestones hit, top contributors, and writes a digest. It lands in your queue Sunday evening. You skim, fix the intro to your voice, approve. It posts to the announcements forum or sends as a DM digest, your choice.
3. Churn re-engagement drafts.
The engagement scorer runs every night and flags members whose activity has dropped below their personal baseline. The re-engagement generator drafts a "we missed you" DM that references the last meaningful thing they posted, or the cohort they were part of, or the topic they were following. You approve the ones that read right. You reject the one for the member you know is on vacation.
4. Pair-up suggestion drafts.
The pair matcher scores members on shared forum subscriptions, recent co-activity, mutual friends, and stated interests. High-score pairs become drafted intro DMs: "you two should meet, here's why." For a community of strangers who would all like each other if they ever talked, this is the function that closes the loop.
5. Forum reply drafts.
The thread scanner watches for threads that have sat unanswered for too long. The reply generator drafts a contextual response from your community manager voice, citing existing wiki entries and prior forum threads where relevant. You approve, you edit, or you let the actual experts in your community take it from here. The point isn't to replace member-to-member discussion. The point is to keep a question from dying.
6. Moderation assist drafts.
When a report lands in the moderation queue, the moderation assist generator reads the report, the flagged content, the member's history, and the rules they may have crossed. It drafts a short summary: what happened, whether it crosses a rule, suggested action, confidence level. You decide. Most reports go from "I have to read 40 posts to make a call" to "I have to make a yes/no decision in 30 seconds."
The approval gate isn't a setting. It's an invariant.
Every AI-generated draft lands in /admin/community-manager with status pending_approval. There is exactly one function in the entire codebase that can flip that status to sent: Dispatcher::executeApproved(). It requires an authenticated admin session and a valid CSRF token. There is no other code path. No background job, no cron, no auto-mode toggle. The AI cannot send on its own.
This is not a configuration choice we trust ourselves to honor. There's a unit test in our CI called HumanInTheLoopTest that walks every PHP file in the source tree and greps for any other write of cm_draft.status='sent'. If a future commit adds one, that test fails the build. It is the canary. The only way to send a draft is for a human admin to click approve.
Why we built it this way: AI quality is good. It is not perfect. Your community's trust in your voice took years to earn and can be lost in one auto-sent embarrassment. Auto-send on community-facing copy is a hazard we declined to ship and made structurally impossible to add later by accident.
BYOK and cost transparency.
You bring your own Anthropic API key. Drop it in /admin/branding. Mobieus does not mark up tokens. You see Anthropic's published prices and you pay Anthropic directly.
You also pick your model per tenant from /admin/config. Claude Haiku 4.5 is $1 per million input tokens and $5 per million output tokens. Sonnet 4.5 is $3/$15. Opus 4.5 is $15/$75. That is a 15x cost spread between Haiku and Opus, and the right answer is not the same for every community.
Order-of-magnitude estimates from our own tenants running all six generators: a 1,000-member community on Haiku is roughly $5 to $15 a month. The same load on Sonnet is roughly $20 to $60. On Opus it is roughly $100 to $300. Your numbers will move with your activity volume; the 30-day spend dashboard in /admin/config shows you exactly where you are, broken down by feature.
Per-feature override is the part most operators end up using. Run Haiku for welcomes and weekly summaries, where the writing is templated and Haiku is fine. Run Sonnet for moderation assist, where reading the thread and getting the call right actually matters. The dropdown lets you mix per feature, so you pay for smart reasoning only where smart reasoning pays you back. The full cost-control story lives in the model selection post.
Why we think this shape is new.
We checked the other community platforms before we said this. Three of them have AI features. The shape is different in each case.
Circle ships AI Agents. They are 24/7 chatbots trained on your posts, comments, courses, and resources, sitting in an AI Inbox. Members converse with the agent. Admins can take over a conversation, and there are Pause AI keywords to hand off sensitive topics. It is a reactive support model, not a proactive engagement model. There are no documented proactive drafts for welcomes, weekly summaries, churn re-engagement, member pairing, or moderation. Circle does not document BYOK or per-model selection on the agent.
Mighty Networks ships AI Cohost as part of its Community Design framework. It helps you design the community on the way in: your ideal member, monthly themes, weekly events, daily questions. It is a strategy assistant, not an ongoing community manager. It is on the Launch Plan at $950 a year. The product page does not describe proactive drafts in members' inboxes, approval queues, BYOK, or model choice.
Skool has not published an AI Community Manager. Their product page leans on simplicity, gamification, and discovery. There is no documented draft-and-approve queue at this writing.
Discourse AI is the most mature AI plugin in the open-source community world. It covers inline assists very well: thread summarization, semantic search, AI helpers for composing posts, AI bot personas, auto-tagging, translation. It is excellent for in-flow help to members. It is not a proactive Community Manager that drafts engagement work and gates on human approval before sending.
If any of these works for the way you run your community, use it. We are not here to dunk. The point is the specific pattern: proactive drafts across six engagement surfaces, sitting in a queue, gated by an admin click, with the cost model under your control. We have not seen another community platform ship that combination. If you find one, tell us and we will update this page.
What this looks like in your week.
Monday morning: the weekly summary draft is in the queue from Sunday night. You tweak the intro to your voice, approve. It posts to the announcements forum. Total time: 2 minutes.
Tuesday: three welcome drafts for the weekend's new members. Two read right; you approve. One mentions a forum you renamed last week; you fix the reference, approve. Total time: 4 minutes.
Wednesday: five churn re-engagement drafts. You approve four. The fifth is for a member who posted yesterday that they are on a two-week trip; you reject that one. Total time: 3 minutes.
Thursday: a forum question that has been quiet for 48 hours has a reply draft attached. The draft cites the wiki page that answers it. You approve. The original poster gets a notification within the hour. Total time: 90 seconds.
Friday: a moderation report lands with an AI assist already attached. The call is "warning, not ban, borderline tone." You read the summary, agree, action it. Total time: 30 seconds instead of the 8 minutes it used to take to read the whole thread.
Most of your community manager work goes from "I have to write this from scratch" to "I have to make a yes-or-no decision in 30 seconds." You stay in the loop. The drafting stops being the thing that doesn't get done.
What we are building next.
We try not to over-promise on roadmap, but a few things are queued and worth naming.
Per-feature model overrides via UI. Today the per-feature override exists in code and ships with sensible defaults. The dropdown to pick Sonnet for moderation assist and Haiku for welcomes is in flight for /admin/config.
More draft kinds. Event-reminder drafts, achievement-grant drafts, and sponsored-thread drafts are the next three on the build list. The generators share a pattern, so adding kinds is bounded work.
Opt-in auto-send for very-high-confidence drafts. Per draft kind, per tenant, off by default. If you set the confidence threshold at 0.95 for welcomes, only welcomes the model is extremely sure of skip the queue. Everything else still waits for a human. We will only ship this with a per-draft-kind toggle, never as a global switch.
Multi-language drafts. Today the generators write in English. Communities that operate primarily in another language will get the model-language hint in tenant settings and the drafts will follow.
Voice memory. The Community Manager learns the patterns of your edits over time so the drafts start to sound like you. Today every draft starts cold; the planned memory layer makes Tuesday's drafts smarter than Monday's because of the edits you made on Monday.
Try it.
The AI Community Manager is available now on Creator Plus at $199/month and Sovereign at $2,500/month. Setup:
- Subscribe to Creator Plus or talk to us about Sovereign.
- Drop your Anthropic API key into
/admin/branding. - Pick your default model in
/admin/config. Haiku is the right place to start. - Watch the queue at
/admin/community-manager. The first welcome draft will land within an hour of your next new member.
If you have not turned the feature on, none of this runs and you are not billed a token. If you have, it runs and you are billed by Anthropic directly. Either way, you see exactly what you spent at /admin/config.
Have a question or a draft kind you wish we shipped? Open a thread at support.mobieus.io/forums/feature-requests. Have a competitor we should have included in the comparison above? Same place. We will update the post.