Comparison — Tusk's Vault vs the alternatives

Honest comparison to the tools TTRPG groups actually use today. Every claim here is verified against the current code — citations link to the source file that backs them. Each tool listed is good at what it's built for; this page is about helping you pick the right one for your table, not picking fights.

🆚 At a glance

Tool	Local-first	Free forever	Source-grounded AI	Discord-native	Voice → chronicle	Open source	No terminal needed	Answers while your machine sleeps
Tusk's Vault	✅	✅	✅ cited	✅	✅ via Tomes	✅ MIT	❌ Node + Git	❌ must be running
NotebookLM (Google)	❌ cloud	✅ (with Google account)	✅ cited	❌	❌	❌	✅	✅
World Anvil	❌ cloud	Free tier; full features paid	❌ no AI	❌	❌	❌	✅	✅
LegendKeeper	❌ cloud	Subscription	❌ no AI	❌	❌	❌	✅	✅
Kanka.io	❌ cloud	Free tier; subscription for advanced	❌ no AI	❌	❌	✅ self-host option	✅	✅
Obsidian + AI plugins	✅	Variable	Plugin-dependent	❌	❌	Plugin-dependent	✅	Sync add-on
ChatGPT / Claude.ai direct	❌ cloud	Limited free tier	❌ hallucinates	❌	❌	❌	✅	✅
D&D Beyond	❌ cloud	Free for character sheets; paid for content	❌ no AI	❌	❌	❌	✅	✅

✅ = strong / built-in. ❌ = absent or significantly worse. The table compresses tradeoffs — the collapsible sections below have the nuance.

Read the last two columns first. Every other column here is one this project was built to win, which is exactly what makes a row of ticks worth nothing on its own. What it loses:

Setup. It wants Node, Git and a terminal before it will do anything.
Always available. It answers from the machine it runs on, so a sleeping PC is a silent bot. (You can still ask from a phone via Discord — as long as that machine is awake.)
Corpus size. The folder source stops including documents past roughly 200 pages (known issues).
Support. There is no team behind it and nobody to escalate to.

Discord is Vault's oldest surface, not its only one: the same answer pipeline also serves Foundry VTT (via a companion module, /tusk <question> in the chat bar) and any MCP client — Claude Code, Claude Desktop, anything that speaks the protocol. See Foundry VTT. "Free forever" means the software; you still pay whichever LLM provider you point it at, or nothing at all on Ollama.

🔍 Vault's claims — verified against the code

Every row of the Vault column is backed by code that exists in this repo:

Claim	Where it's enforced
Local-first	Default bind is `127.0.0.1` — see `host-origin-guard.ts`, which refuses requests whose `Host` is not a loopback literal.
Source-grounded	Canonical seven-rule block in `prompt/system.ts → CANONICAL_RULES_BLOCK` is force-appended whenever the chosen system prompt doesn't already contain it — citation discipline is structurally guaranteed, not configurable away.
Off-corpus refusal	Rule 3 requires the bot to emit the canonical refusal phrase verbatim; the lore-gap detector picks it up via `LORE_GAP_TRIGGER_FRAGMENT` and routes the question to the dashboard's Lore Gaps tab.
Discord-native	Bot client in `src/server/discord/`, message handling in `src/server/surfaces/discord.ts`.
Three ask surfaces, one pipeline	`SurfaceId` is `discord \| foundry \| mcp` (`chat/types.ts`); all three call the same `chat/ask.ts`, so persona, corpus and citation discipline are identical wherever the question came from.
Encrypted keys	AES-256-GCM, machine-bound, in the platform config dir — see `keys/crypto.ts`. Never plaintext on disk.
MIT open-source	LICENSE.
One outbound network call per answer	The LLM adapter — `openrouter.ts` or `gemini.ts`, Ollama over loopback, or a local Claude Code process that talks to Anthropic on your behalf. Discord adds its gateway connection when that surface is on; Foundry and MCP add none. No telemetry, no analytics, no phone-home.
Nothing to install inside the app	Local models and personas are ordinary features, not modules. There is no plugin system to learn and nothing to enable.

⚰️ What happens if this project dies

Worth asking of anything you are about to trust with years of worldbuilding, and it is the fair question to put to a project maintained by one person.

Your source material is untouched. Vault reads the folder you pointed it at. It does not convert your documents, import them into a database, or move them — there is no database, and Tusks-Lore/ after a year of use contains the same files you put there. Delete Vault and the folder is exactly what it was.

There is no export step, because nothing was captured. Everything Vault adds sits beside your documents as plain, pretty-printed JSON you can open in any editor:

File	What it holds
`clarifications.json`	The rulings you recorded, question and answer as text
`lore_gaps.json`	Questions your notes did not cover
`clarifications.embeddings.json`	Base64 vectors used for matching — derived data, regenerable

An export button exists in hosted tools because the tool is holding something of yours. Nothing here is holding anything.

Nothing switches off, because nothing is running anywhere else. There are no servers, no account, no licence check and no phone-home, so there is no lever anyone can pull to stop an install that already works. It keeps working whether or not the project does.

The code is MIT. Fork it, patch it, hand it to whoever wants to keep it going, or lift the one part you liked. LICENSE.

What you would actually lose

Not nothing, and the honest version matters more than the reassuring one:

Fixes stop. Bugs on the known issues list stay there.
Nothing adapts around it. If Discord changes its gateway, or a provider changes its API or retires a model, an unmaintained Vault breaks at that surface and stays broken until someone patches it. This is the realistic way it dies — not deletion, but drift.
Dependencies rot. The security advisories that ship in any Node project stop being triaged.

So the risk is real, but it is bounded, and it is a different shape from the one hosted tools carry. There, the failure mode is your material becoming inaccessible on somebody else's schedule. Here, the failure mode is a tool that stops improving next to a folder of files that never depended on it.

🧭 Tool-by-tool — the nuance

One caveat before the list: everything said about Vault below is checked against this repository's code, but everything said about the other tools is a snapshot of products that are actively developed and priced by other people. Several of them are adding AI features. Check the current state of anything you are about to pay for rather than taking this page's word for it.

NotebookLM (Google)

This is the closest comparison, so it gets the longest entry — and the concession first.

What it's great at, and better at than this: cloud-hosted retrieval over documents you upload, with exact source citations and a genuinely excellent interface. Free with a Google account, nothing to install, and built by a large team with a budget — it will keep improving faster than a one-person project can.

If what you want is to interrogate a pile of documents carefully and on your own, use NotebookLM. This is not trying to beat it at that.

What it does not attempt. The distinction worth drawing is not quality, it is purpose. NotebookLM is built for a reader alone with their sources. Almost everything below is something a research notebook has no reason to have an opinion about:

It is a tab you visit, not a place answers appear. Vault answers in your campaign Discord and in Foundry's chat bar, where the argument is already happening and everyone at the table sees the reply. That is the difference between settling a question and looking one up. (A per-world access mode decides whether players may ask at all, and whether answers are whispered or posted — see Foundry VTT.)
It has one voice. Vault ships eight personas, from a scholarly archivist to a foul-mouthed sellsword, and the citation rules survive all of them. A table's tone is half of why anyone enjoys asking.
It will not speculate, deliberately. "Who is most likely to start a tavern fight" has no answer in anybody's notes. Vault's Speculative Mode reasons from established character traits, cites the evidence it used, and tags the conclusion [speculation] so a guess is never mistaken later for something the GM wrote. NotebookLM's grounding is a design commitment, and a correct one for research; it just rules out the questions a table most enjoys.
It sanitises grim material. Campaign fiction inherits its genre — war, cruelty, horror, grief. Vault's prompt states plainly what the corpus is (authored fiction, the asker is one of its authors, the task is recitation) and recounts a scene as written. A summariser that softens the author's own writing back to them is failing at the job.
A GM ruling is not a first-class thing. Vault's clarifications persist, embed, and get re-applied to differently-worded questions later, outranking the documents when they disagree. NotebookLM has no concept of "the GM overrides the source".
Your documents leave your machine. Uploads live on Google's infrastructure. Vault keeps the corpus and the index local, sends only matched passages to the model you chose, and can run entirely offline against a local model.
You cannot change it. Vault is MIT, forkable, and lets you pick the model — or swap it mid-campaign.

Pick NotebookLM if you want a polished cloud research notebook, you are mostly working alone, and you are content for Google to hold your campaign.

Pick Vault if the answer needs to land where your group can see it, in a voice they picked, about material a cloud summariser would flinch at — and you would rather the whole thing ran on your own machine.

Honestly: both. They are not substitutes. Nothing stops you researching a setting in NotebookLM and running Vault at the table.

World Anvil

What it's great at: worldbuilding as a hypertext wiki. Beautiful presentation. Excellent for setting publication.