Our taste here

A lot of AI in community software is decorative: a chat bubble that knows a little less than the docs page, a "summarize this" button on a four-message thread, an autoresponder that confidently picks the wrong forum to move you to. We did not want any of that. Each feature in this release exists because a real operator told us they spend time on it manually. Each one ships off by default. Each one runs against your own Anthropic key, billed at cost to you, with no Mobieus markup. And each one caches results so the same input is not re-billed every time someone refreshes.

What shipped

1. Long-thread summaries with key points

Long threads are useful for the people in them and hostile to everyone who arrives late. On any thread above a configurable post count, members see a "Summary" panel above the first post: a four-to-six sentence overview, then a short bullet list of key points. Click expand for the full prompt context. The summary is cached per thread and regenerates when the thread grows past a delta.

2. Suggested tags when composing

When a member writes a new thread, the composer can suggest up to five tags based on the title and body. The model returns JSON, the UI shows them as chips, the author clicks the ones that fit. Tags are cached by a SHA-256 of the content, so saving a draft and reopening does not re-bill the API.

3. Search synthesis on the results page

When a member searches across forums, files, threads, and the wiki, the top of the results page now has a "Synthesize" button that turns the top results into a single short answer with citations back to the source rows. Best for the question "did anyone already answer this." Results are cached for six hours per query hash so a busy moderator-of-the-day does not run up the bill.

4. One-click explain on moderator reports

On the moderation queue, every report has an "Explain" button. The model reads the reported post, the report reason, and the recent context, then writes a two-paragraph explanation of what the report likely means and what the policy options look like. Useful for a new moderator who is reading their first 50 reports; useful for the senior moderator who wants a sanity check before issuing a ban.

How we gated this

Three layers, in order:

  1. Global feature flag. An admin sets AI_FEATURES_ENABLED in /admin/config. Off by default. Off means the controllers refuse the request and the UI buttons do not render.
  2. Your Anthropic key. The platform never includes our key. You bring your own at /admin/config?tab=environment. The platform calls Anthropic directly from your tenant, billed to your account.
  3. Per-feature caches. Thread summaries cache per thread id (regen on post-count delta). Tag suggestions cache per content hash. Search synthesis caches per query hash with a six-hour TTL. Report explanations cache per report id (one-shot).

This means the cost shape is bounded. You decide what to enable, you see your own bill, you do not pay twice for the same input.

What we deliberately did not ship

  • A general-purpose chatbot. Plenty of products have one. We did not want to ship the seventh-best chatbot on the internet.
  • Automatic moderator actions. The Explain button writes a paragraph; the human decides.
  • Auto-replies in threads. We will not turn your community into a place where the bot answers before the regulars do.
  • Tag auto-application. Suggested tags are suggestions; the author clicks the ones they accept.

What you can do today

  • Enable AI_FEATURES_ENABLED in /admin/config.
  • Paste your Anthropic key into /admin/config?tab=environment.
  • Open a long thread and try the Summary panel.
  • Compose a new thread and click "Suggest tags".
  • Search for something and click "Synthesize" on the results page.
  • Open a report in /admin/moderation and click "Explain".

What is next

The catalog grows where operators say it should. The candidates we hear most: a "rewrite this for clarity" assist on long replies, an "auto-categorize" tool for the marketplace, a digest-generation assist for the email builder. If one of these would change your week, tell us.