Questions & Answers

Hubbards.AI is currently free to use. We carry the cost of the underlying AI models on your behalf.

Because those LLM calls aren't free for us, each free account has a daily and monthly query budget. The budget is measured in LLM-cost, not in question count — so the model you pick directly shapes how much you can ask:

• Cheaper models (at the top of the dropdown, e.g. Gemma) consume very little of your budget. Lots of room for many questions per day.
• The premium tier in the free phase is Claude Haiku. It gives richer, deeper answers than the cheap models, but each query uses more of the daily budget — so fewer questions before the limit kicks in. (Claude Sonnet and Claude Opus are not in the free-tier dropdown right now; they'll come back with the sponsor plans.)

A good workflow: scout broadly with a cheap model, then re-ask the interesting ones with a bigger model for a more thorough synthesis.

Sponsor plans are coming. They will remove the daily and monthly limits in return for helping cover the running cost. Want to be notified when they go live? Subscribe to our Telegram channel — we'll announce it there first.

All models see the same LRH source chunks — what differs is how deeply they reason over them and how fluent the synthesis reads.

• Llama 3.1 8B / Gemma 3 12B — fast, very cheap. Great default for casual look-ups, definitions, single-passage questions. Hundreds of queries per day fit in the free budget.
• Gemini 2.0 Flash / Llama 3.3 70B / GPT-4o-mini / DeepSeek V3 — middle tier. Solid reasoning across multiple sources. Good for "what does LRH say about X across the years" type questions. Roughly 80–170 queries per day in the free budget.
• Claude Haiku 4.5 (Anthropic) — premium tier in the current free phase. Best for nuanced, multi-layered queries that need to weigh competing passages or trace a concept's development across the timeline. About a dozen thorough questions per day in the free budget.

Citation quality is the same across all models: each answer is grounded in the same verbatim corpus chunks.

Note: Claude Sonnet and Claude Opus are currently disabled in the free tier. They re-appear once sponsor plans launch.

Two sliders/checkboxes in the right sidebar shape what the AI gets to see.

Treffer (k) — number of source chunks

A slider from 3 to 20, default 8. This is how many of the most relevant verbatim passages the search hands to the AI before it answers.

• Keep it at 8 for typical look-ups, definitions, and "where did LRH say X?" questions. 8 passages is usually plenty.
• Raise to 12–20 when: the question spans many sources ("How did LRH's view on auditor presence evolve from 1952 to 1970?"); you tried with 8 and the answer feels incomplete; you suspect a specific HCOB exists but the search didn't surface it in the first 8 hits.
• Lower to 3–5 only if you want a sharply focused single-quote answer and you trust the top result; the AI has less to weigh, the answer comes out leaner.

Budget impact: more chunks = more input tokens to the AI. On Gemma or Llama 8B, going from k=8 to k=20 is essentially free (fractions of a cent). On Claude Haiku it's noticeable — a k=20 Haiku query roughly doubles the cost of a k=8 one. Rule of thumb: raise k freely on cheap models, raise it on Haiku only when you really need the breadth.

"Auch Post-1972-Material einschließen"

Off by default. The search is restricted to LRH's original canon (up to 8 June 1972) — the body of work before the Church of Scientology began systematically rewriting his bulletins, policies and books.

• Leave it off if you want answers grounded in what LRH actually wrote and said in his own lifetime.
• Turn it on if you want to compare original vs. post-1972 revised versions, look up something specifically from a Ron's Journal or later issue, or you suspect the answer only exists in post-72 material.

Post-72 hits are not hidden because they are wrong — many are not — but because they may be modifications that don't reflect LRH's original intent. With the toggle on, the citation still tells you the exact date, so you always know which material you're looking at.

No budget impact from this toggle — it only changes which passages are candidates, not how many.

The OT III control is covered in its own Q&A below.

There's a fair-use throttle of roughly 30 requests per minute (with short bursts allowed) to keep the service responsive for everyone. That's well above normal use — even a thorough multi-step research session won't hit it.

The cost-based daily/monthly budget (see the cost Q&A above) is what bounds free-tier consumption, not the per-minute throttle.

Around tens of thousands of indexed passages drawn from the body of L. Ron Hubbard's recorded and written work:

• HCOBs — Hubbard Communications Office Bulletins (the technical line)
• HCOPLs — Policy Letters (the administrative line)
• Full lecture transcripts (Congresses, ACCs, Saint Hill lectures, etc.)
• Books — Dianetics, the Basics, axioms, etymology
• Ron's Journals, executive directives, other issues

Each chunk is dated and tagged by issue type so you can cite the exact HCOB, PL, lecture or book the answer came from.

The search is semantic (meaning-based), not keyword-only — paraphrase the concept in your own words and the relevant LRH passage surfaces.

Notice something missing? Drop us a note through the contact form in the app and let us know what's not there. Chances are we'll add it.

The corpus is in LRH's original wording — almost entirely English. The semantic search works best when your query is in the same language as the source, so English queries match best.

You can ask in German, French, Italian, Russian, etc. — the AI model will translate internally and try to match. Quality is usually still good for well-known concepts, but for half-remembered references it pays off to drop in any English keyword you remember (a term, a date, an HCOB number).

The answer comes back in the language you asked in, with verbatim English source quotes preserved.

Tip on model choice for non-English queries: larger models handle other languages noticeably better than the cheapest ones. If you ask in German, French, Italian, Russian, etc., it pays off to step up from Llama 8B or Gemma to a mid-tier model (Gemini 2.0 Flash, GPT-4o-mini, DeepSeek V3, Llama 70B) — or to Claude Haiku for the most fluent translation and reasoning. The tiny models can stumble over idioms and produce awkward German; the larger ones do not.

Your questions are not used to train any AI model — not by us, and not by our LLM providers (Anthropic and Google explicitly exclude API traffic from training).

We log basic usage data (per-account query count, model picked, timestamps) to operate the service, monitor abuse and bill ourselves correctly against the LLM providers. Query content is not retained long-term beyond that.

The site uses privacy-friendly analytics (Matomo, self-hosted, cookieless, IP-anonymised, DNT-respected) for traffic counts only.

Confidential material (OT III references, related BT / wall-of-fire content) is filtered out by default as a precaution against restimulation for unprepared readers. In the web app there is a sidebar checkbox to override this. Via the MCP tool, pass include_ot3=True in the search_lrh call.

Yes — any MCP client that speaks the SSE transport works.

Claude Desktop: edit claude_desktop_config.json and add:

{
  "mcpServers": {
    "hubbards-ai": {
      "transport": "sse",
      "url": "https://hubbards.ai/sse",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Cursor, Cline, Continue, others: check the MCP-server config section of your client. Configure SSE transport with URL https://hubbards.ai/sse and an Authorization: Bearer YOUR_TOKEN header. That's it.

Go back to auth.hubbards.ai/if/user/ → "Tokens and App passwords", find the token you want to kill, delete it.

Server-side token cache expires within 5 minutes, after which all clients using that token get rejected. Create a fresh token the same way if you just want to rotate it.

app.hubbards.ai is a full chat UI: you ask a question, our hosted LLM (Claude, Gemma, your pick) reasons over the retrieved chunks and gives you a synthesized answer with citations. Convenient, no setup, daily budget.

The MCP server is the same retrieval engine, exposed as a tool for your own AI client. It returns chunks only; your AI model does the synthesis. You bring your own model + key, you control the cost, you keep the conversation in your own client.

Use the web app for casual queries. Use MCP when you want LRH lookups inside a longer Claude Code / Claude Desktop conversation.

MCP is an alternative way to access Hubbards.AI — instead of using the web app in your browser, you wire Hubbards.AI directly into a local AI tool you already use (Claude Code, Claude Desktop, Cursor, Cline, etc.) and ask it questions from there.

Why bother — what is the advantage?

• Unlimited queries. You bring your own AI model + key, so the daily budget on hubbards.ai/app does not apply. Ask as much as you want.
• No per-query cost from us. You only pay your AI provider for the model calls you make.
• Stays inside your existing workflow. If you already live in Claude Code or Claude Desktop, you don't switch tabs — Hubbards.AI shows up as one of your tools and you can chain it with anything else (files, web, your codebase, your notes).
• Verbatim, citable. Same corpus as the web app (~49 000 indexed passages, every chunk tagged with HCOB / HCO PL / lecture / book + date + section). Your local AI reasons over the retrieved chunks and cites the exact sources.

Roughly: the web app is for casual look-ups; MCP is for serious research sessions or anything where you want volume.

How to set it up — three steps:

1. Sign up at Hubbards.AI. Go to app.hubbards.ai and log in (free, passwordless e-mail login).
2. Create an API token. Once logged in, open auth.hubbards.ai/if/user/, click "Tokens and App passwords" → "Create Token". Copy the token immediately — you only see it once.
3. Register the server in your MCP client. In Claude Code, run in your terminal:

   claude mcp add --transport sse hubbards-ai https://hubbards.ai/sse \
     --header "Authorization: Bearer YOUR_TOKEN"

   Replace YOUR_TOKEN with the token from step 2.

Verify the connection:

claude mcp list

You should see hubbards-ai with status connected.

What you get on the tool side:

A tool called search_lrh(query, k=8) returning the top-k most relevant verbatim chunks from the LRH corpus. Each chunk comes with full source data: title, date, issue type (HCOB, HCO PL, lecture, book, congress), section, and the original text.

Example use inside any Claude Code session:

Use search_lrh to find what LRH said about
missed withholds in 1962.

Claude Code will call the tool, receive verbatim chunks, then reason and cite the exact sources for its answer.

OT III material is filtered out by default — see the confidential-material Q&A below.