Ask your campaign anything

The answer that sounds right is the dangerous one

My campaign notes were spread across six documents, two of which contradicted the other four. Players asked me the same questions between sessions — who was that merchant, what did we actually agree to, whose brother is the king — and I answered from memory, badly, and sometimes differently than last time.

The obvious fix is to ask an AI, and it fails in a specific way: a general chatbot handed your setting answers every question confidently, including the ones your documents say nothing about.

So the whole design is one idea: it may only tell you what your documents say, and it must show you where it got it. Answer a logged gap once and that answer is reused from then on.

Nobody needs this until the evening a player asks something specific and four people remember it four different ways. Then it is the only thing you want.

The best thing it does is admit defeat

Any assistant can answer a question your documents cover. The one that decides whether you can trust it at the table is the question they don't — and the useful behaviour there is to stop, say so, and remember that you were asked.

A ruling recorded this way is cited like any other source, so six months later nobody has to remember whether it was decided or imagined. What people actually ask it →

Real questions, real citations, and one refusal, against a corpus written to be published. See the answers in full →

Why not use something that already exists?

Often you should. Most people arrive here having already pasted their notes into ChatGPT — good software, and for a question your documents genuinely answer it will usually answer correctly. The trouble is the question they don't.

NotebookLM is the closest comparison, and on its own ground it is better than this. Google's budget, excellent retrieval, exact citations, nothing to install. If you are working alone through a pile of documents, use it. The difference is not quality — it is what the two things are built for.

The second column is also this project. The Foundry module answers on its own, out of the journal entries in the folders you choose, with nothing else installed — that is Lite, and for a campaign that already lives in Foundry it may be the whole answer. It sits in the grid because the honest way to show what the full application adds is to put the two beside each other and let the rows say it.

Which is the honest shape of it. This is not a better NotebookLM — it is a worse one, pointed somewhere NotebookLM is not pointed. Keep theirs for research; use this when the question is being shouted across a table. Campaign wikis and your own GM are not in that grid at all, because neither is a thing you ask — both, and the case for using them instead, are in the full comparison.

Nine of these rows go this project’s way. Four do not. You would have found the four out on your own soon enough.

Four stages, and one of them is the point

A question goes in one end. Stage two is the one you cannot configure away, and the reason the answer at the far end can be trusted.

1. 1

   Find what bears on the question

   Reads a folder on your disk — PDF, DOCX, Markdown, plain text — or an Obsidian vault, read-only, nothing moved. Passage-level retrieval then picks out the parts that matter, so a large campaign is not a large bill.

2. 2

   Force the rules on top

   Cite every claim; refuse rather than invent. The block is force-appended to every prompt, so editing the system prompt — or picking a persona — cannot remove it.

   The reason this exists

3. 3

   Ask your model

   Gemini, ~400 models through OpenRouter, a Claude subscription, or one running on your own machine. Swappable mid-campaign.

4. 4

   Check what came back

   A refusal is labelled rather than served as content in the archivist's voice — and the question it could not answer lands in your dashboard as a lore gap.

The corpus travels with every question, so every cloud request pins a zero-data-retention routing floor. How the retrieval pipeline works →

Not every question is a lookup

Most of what gets asked at a table is not "when did Dunmar fall". It is an argument two players are having, or a joke someone wants settled with evidence. Those work — and they work for the same reason the serious ones do. It can only answer from what you wrote down, and it shows you where.

And it doesn't have to sound like a librarian

A persona replaces the whole voice, and the citation rules survive it. Same question, same corpus, same sources — two of the eight that ship:

And you decide whether it's allowed to guess

"Who would lose a duel to a chicken" is not in anybody's notes, so by default it is refused like any other question the documents don't answer. Turn on Speculative Mode and it will reason from what the chronicle establishes about each character — then cite the evidence it used and tag the conclusion, so a guess can never be mistaken later for something you wrote down.

Three ways in, one answer

A lore assistant nobody opens is a lore assistant nobody uses, so it answers where the question already gets asked.

The third route needs no setup screen at all: any MCP client — Claude Code, Claude Desktop, anything speaking the protocol — can pair and ask directly. What people actually ask it →

And it doesn't have to be a campaign. Nothing in the machinery knows what a campaign is — it indexes a folder, cites what it finds, and refuses what isn't there. Point it at a handbook or a wiki export and you get the same bot.

The prompts, the citation styles and the refusal wording all assume a GM and a table. Everything else is a happy accident.

A question, not a month

Tusk's Vault is free. The only money involved is whatever your chosen model charges for the tokens, and a single question is a very small number of them — one retrieval-augmented request, not a generation pipeline. Gemini's free tier covers most home games outright, and a local model costs nothing at all.

The full cost breakdown →

You're handing this your whole campaign

Years of somebody's worldbuilding, and often things said at a table that were never meant to leave it. So it's worth being exact about where everything goes. Almost all of it stays on the machine you installed it on.

• No account, no sign-up, no licence check.
• No telemetry, no analytics, nothing phoning home. Ever.
• Keys encrypted at rest and tied to the machine — copying the file elsewhere gains nothing.
• Cloud requests pin a zero-data-retention floor, because the corpus travels with the question.
• MIT licensed, every line readable, including the setup scripts.

One person, one Windows PC, no support desk — which is why the feedback button sits next to the coffee one in the app. Telling me what broke gets it seen; it does not promise the fix sticks. The unvarnished version is here.

So what happens if I stop?

Worth asking of anything you hand years of worldbuilding. The answer splits cleanly.

And who should be allowed to ask?

A different question from where the data goes, and the one more likely to matter at a table. Anyone who can ask reaches the whole corpus, because retrieval follows the question — ask about a mystery and you get the notes about that mystery. That is what makes it useful, and it is worth knowing before you point it at a group.

Players will also try talking it into simply reciting what it was handed. Some of them will manage it. What comes out is your own lore — not your keys, your files, or anything else on the machine. The archivist has no filesystem, no shell and no network of its own, and that is tested rather than hoped for. So the thing to plan around is spoilers and timing, not a breach.

Which makes this a tool for a private Discord or a Foundry table of trusted players rather than a public server. Player questions are off by default on Foundry; on Discord the bot only reads the channels you put it in. What a prompt injection can and cannot do spells out both halves.

The ones people actually ask