AI Helper turns Claude — or any model you prefer — into a member of your support team. Mention it, or point it at your support tags, and it replies in seconds using your own forum's posts as context. That means answers about your product, your rules, and your history — not generic AI filler.

## ✨ What it does

- 🤖 Answers automatically — replies to @mentions and to new posts in the tags you choose.

- 🧠 Grounded in your forum (RAG) — it searches your existing posts and answers from them, so replies stay accurate and on-topic.

- 🔌 Bring any model — native Anthropic (Claude), or any OpenAI-compatible provider (OpenAI, OpenRouter, Groq, Azure, even local models).

- 💸 Exact cost tracking — an admin dashboard shows spend vs. budget with a daily graph. Numbers come straight from the API, not estimates.

- 🛑 Hard budget cap — set a monthly limit and the bot pauses itself when it's reached. No surprise bills.

- ✍️ Auto or Draft mode — post answers directly, or hold them for a moderator to approve and edit first.

- 📝 One-click summaries — any member can get a clean TL;DR of a long thread.

## 🧠 "Grounded in your forum" — what that means

On its own, an AI model has no idea what's inside your community. AI Helper fixes that with retrieval: before answering, it finds your most relevant past posts and feeds them to the model. Two modes:

- Keyword (FULLTEXT) — works out of the box, no extra key.

- Semantic (embeddings) — optional upgrade that matches by meaning (so "can't log in" finds "password won't work"). Costs pennies.

## 🚀 Getting started

1. Install, then open Admin → AI Helper.

2. Add your API key, pick a model, and set a monthly budget.

3. Choose Auto or Draft, select your support tags, and switch it on.

Bring your own API key — it stays on your server, and the bot only ever sees content the asker can already view.

This is a 💸 Premium Extension that must be purchased on Floxum. Cost is a one time charge of $5.00 since you also have to pay for your API tokens as well.

ai-helper 0.1.0

🎉 First release of AI Helper — an AI support assistant for Flarum 2 that answers your members, grounded in your own forum.

Highlights

• 🤖 Auto-answers @mentions and new posts in the tags you choose (visual tag picker in admin)
• 🧠 Retrieval-grounded (RAG) — keyword FULLTEXT (keyless) or semantic vector embeddings
• 🔌 Multi-LLM — Anthropic (Claude) or any OpenAI-compatible provider (OpenAI, OpenRouter, Groq, Azure, local)
• 💸 Exact cost tracking — admin dashboard with spend-vs-budget bar, daily graph, and a hard monthly budget cap
• ✍️ Auto or Draft (moderator-approval) reply modes
• 📝 One-click thread summaries

See the README for setup. Bring your own API key — it stays on your server.

ai-helper 0.2.0

Bug & suggestion triage 🐛💡

Claude_Bot can now capture bugs and feature requests from your forum and turn them into a backlog you act on later.

• 🔎 Classifies posts it answers as bug / suggestion / question
• 🗂️ Records bugs & suggestions in a new admin Triage queue
• 🐙 Mirrors them to GitHub Issues in the right repo (tag → repo routing), with a link back to the thread
• ⚡ Hybrid: auto-files high-confidence bugs; holds the rest for one-click approval
• 🔁 Dedupes so repeat reports don't create duplicate issues

Configure under Admin → AI Helper: enable, GitHub token, default repo, tag → repo map, auto-file confidence.

ai-helper 0.2.1

Fixes

• 🏷️ Watched-tags picker now shows secondary & child tags, not just primary ones. The admin only pre-loads primary tags, so the picker was skipping the rest — it now always loads the full tag set and groups child tags under their primaries.

(Tag matching for both auto-replies and triage routing already worked on every tag attached to a discussion — this just makes them all selectable in the picker.)

ai-helper 0.3.0

Auto-detect repositories 🪄

Setting up triage routing is now one click. The new Auto-detect repositories button (Admin → AI Helper, next to the tag→repo map) lists the GitHub repos your token owns, matches them to your forum's tags (exact names + flarum-/word-order aliases), and fills the map for you to review and Save.

• 🔎 Matches forum tag slugs → owner/repo
• 🧩 Handles aliases (e.g. flarum-seo → seo, facebook-flarum-post → flarum-facebook-post)
• 🔁 Re-run anytime to pick up new extensions

Save your GitHub token first, then click Auto-detect.

ai-helper 0.4.0

Added

• DM triage. Admins can direct-message the bot — e.g. paste audit results — and every bug/suggestion is logged to the triage queue + GitHub, with the bot replying in the conversation. One message can hold many findings (an audit dump becomes many tracked items), each routed to the right repo (owner/repo hint → tag map → default). Requires flarum/messages; turn on Capture from direct messages in settings. Admin senders only.

ai-helper 0.5.0

What's Changed

• ⚙️ Run the bot without a persistent worker — new opt-in setting "Process the queue on the scheduler". For hosts that run php flarum schedule:run every minute but can't keep a queue:work daemon alive (shared hosting), the scheduler now drains the queue each minute so bot replies, DM triage and notifications are processed. Leave it off if you already run a dedicated queue:work.

ai-helper 0.6.0

What's Changed

• 💬 Direct-message chat — new setting "Reply to direct messages (chat)". DMing the bot now gets a real conversational reply (answers questions, acknowledges instructions), like @mentioning it but private. Admin DMs still get bug/suggestion capture on top. Budget-capped and loop-safe.

ai-helper 0.7.0

What's Changed

• feat: index extension docs (README + composer) into the knowledge base
• feat: index extension locale strings; fix README truncation

ai-helper 0.8.0

What's Changed

• feat: auto-refresh extension-doc index (no shell needed on managed hosts)

ai-helper 0.8.1

Floxum audit hardening — all 10 findings (Q:63 → ≥95), validated end-to-end on a live Flarum 2 install (Redis queue + realtime + MySQL + a real LLM round-trip).

• DB portability: MATCH…AGAINST retrieval now driver-guarded with a portable LIKE fallback (PostgreSQL/SQLite); the content FULLTEXT index is created only on MySQL; FIELD() ordering → portable CASE.
• Production: MaintainIndex no longer embeds inline on the request — it pushes a queued IndexPost job (verified running async on a worker); N+1s in the drafts + triage controllers batch-loaded; a per-user fixed-window rate limit on the Summarize endpoint (10 / 5 min).
• Structure: duplicated chunk()/embed() extracted to a shared ChunkEmbed trait; a new Model\Knowledge Eloquent model replaces raw query-builder access; dead build-phase scaffolding removed from extend.php.