AI Dungeon alternatives: which AI RPG actually plays by rules — and in your language?

AI Dungeon is English-only. That isn't a complaint, it's the company's own answer in its own help centre — and if English is the language you play in, it sounds like the question is settled. It isn't. Every large language model writes fluent English in 2026, so fluency stopped being a differentiator; what separates these platforms now is whether there is any product underneath the prose — who throws the dice, where your hit points are written down, and what happens when you want to bring a friend who plays in German. This page weighs six platforms on those two axes, plays the same opening scene on four of them, and says plainly where the competition beats us.

Is AI Dungeon English-only? The short answer#

Yes. AI Dungeon officially supports English and nothing else, and that isn't an outsider's reading — it's the company's own reply to the language question in its help centre, which opens with "AI Dungeon is currently only available in English." The same page gives the reasons: the model's training data, and a small team that can't keep an interface alive in a dozen languages. It then notes that some players have managed anyway, by writing directly in another language or running the whole thing through translation software.

So if you type in French or Japanese, you will usually get French or Japanese back. That is the language model's doing, not the product's. The interface, the billing screen, the error messages, the scenario library and the community all stay in English, and a long scene will often drift back to English on its own halfway through. For an English-speaking player none of that is a problem. It becomes one the moment you want a table where everyone reads the narration comfortably.

Writing a language vs. being built in it#

You can't settle whether a platform "speaks a language" with a single question, because there are three tiers and all three put readable text in the chat box. Telling them apart means looking outside the box.

The three tiers

• Tier 1 — one language only. Interface and content are English. Write in another and the model improvises; the product doesn't. AI Dungeon and Friends & Fables sit here.
• Tier 2 — translated. The interface has been bulk-translated, and the narration is thought in English and then spoken in other words. The sentences are correct but the voice is imported — the tone of a manual, not a storyteller. Most multilingual story platforms sit here.
• Tier 3 — written natively. Narration is generated directly in the chosen language, interface and system strings were written in it, errors and support come in it too. Very few platforms sit here.

Five checks tell you which tier you're looking at, and none takes a minute. First, break something on purpose: trigger an invalid action and read the language of the error — the translation layer collapses there before anywhere else. Second, open the billing or account screen; marketing pages get localised, invoices almost never do. Third, throw an idiom into the narration and watch not whether the model understands yours, but whether it ever coins one of its own unprompted. Fourth, check voice consistency: if the Game Master slips between present and past tense, or between second and third person, every few turns, there is no style plan behind it. Fifth, look for strings that don't fit their boxes — clipped buttons, "1 items", a date rendered 08/20/2026 on a page whose readers write 20/08/2026. Those small seams tell you the interface was fitted afterwards.

AI RPGs compared: language and rules#

I compared six platforms across ten headings: interface language (menus, errors, billing), narration quality, localised content and support, rules engine, who throws the dice, persistent world state (hit points, inventory, gold, quest step), multiplayer, art and voice, free trial, and the single thing it does best. Where a cell fell outside what I actually tested, I marked it untested rather than inventing an answer.

AI Dungeon — the freedom champion, in one language

• Interface language: English — menus, errors and billing screens included.
• Narration: excellent in English. Write in anything else and the model improvises; there's no official support, and long scenes drift back.
• Localised content and support: none. Help centre, scenario library and community are English.
• Rules engine: none; free narration.
• Who throws the dice: dice aren't part of the product. Bolt them onto a scenario yourself and the number is produced by the same model that narrates it.
• Persistent world state: memory and note fields exist, but there's no numeric ledger — hit points and inventory live inside the prose.
• Multiplayer: yes.
• Art and voice: scene image generation.
• Free trial: yes.
• Best at: limitless freedom, and the biggest scenario archive in the category.

This card deserves to be written generously, because this is the product that created the category. Without Latitude shipping AI Dungeon in 2019, most of the platforms compared on this page wouldn't exist. Its strength is still exactly where it always was: no genre boundary, no tone boundary, no definition of what counts as a "game". You can go from a dragon hunt to a space station to an absurd scene in a kitchen in a single sentence and the model won't object. Nobody else has the library of worlds and scenarios that has piled up there over the years. Its weakness is the price of that same flexibility: no rules, no ledger, no referee. And the language answer cuts both ways — perfect if English is where you live, a dead end the day your table isn't.

Friends & Fables — takes the rules most seriously

• Interface language: English.
• Narration: English-centred; other languages aren't officially offered.
• Localised content and support: none.
• Rules engine: yes — built on D&D 5e and it genuinely applies the mechanics.
• Who throws the dice: the system handles rolls; they aren't left to the narrator's mood.
• Persistent world state: character sheet, inventory and level progression are tracked.
• Multiplayer: yes, played as a party.
• Art and voice: the best tactical map in the category; combat really does happen on a grid.
• Free trial: yes.
• Best at: rules depth and visualising combat.

If you've spent years at a physical table, this is the product that makes us look thinnest on the rules axis. Initiative, distance, action economy and character progression are taken seriously, and the tactical map experience is clearly ahead of ours. Its limit is reach rather than quality: it understands the rules, but it only speaks one language, so a mixed-language group has nowhere to sit.

Multilingual story platforms (aiga_ and similar) — many languages, no rules

• Interface language: 16 languages. Real work has gone in on this side.
• Narration: available in all of them, but it reads translated: correct sentences, imported voice.
• Localised content and support: yes — they publish localised blog and comparison pages.
• Rules engine: none. On their own comparison pages they describe the product as narrative-focused rather than a strict tabletop rules engine.
• Who throws the dice: there's no separate dice authority; the model that narrates the outcome decides it.
• Persistent world state: narrative memory, no numeric ledger.
• Multiplayer: not the centre of the product.
• Art and voice: outside my test scope.
• Free trial: yes — three turns a day on the free tier.
• Best at: language breadth; one of the few that can honestly answer yes to "can I play in my language?"

Skipping this section would have made the page dishonest: claiming we're the only platform that plays outside English would simply be false. aiga.io ships a localised interface in sixteen languages, publishes content in them, and lets you play three turns a day for free. On the language axis they're attacking the same problem we are, and they've solved a real part of it. The difference is tier: the localised text reads like it came through English on the way, which is a choice about cost, not a bug. The second and larger difference is the rules axis, and we don't have to assert it — they do. The product is narrative-first and doesn't claim to be a rules engine. Dice, hit points and inventory live inside the story. That's a deliberate design, not a flaw — but if the feeling you're chasing is "I rolled badly, so I failed", this isn't where you'll find it.

Telling a chat assistant "be my Game Master" — free, and why it falls apart

• Interface language: whatever you speak.
• Narration: superb. The most fluent prose on this list can come out of here.
• Localised content and support: yes.
• Rules engine: whatever you build; it lives as long as the model keeps applying it and ends the moment it forgets.
• Who throws the dice: the model. The thing producing the number and the thing writing the story are the same thing.
• Persistent world state: none. Memory holds preferences, not a ledger.
• Multiplayer: no.
• Art and voice: image generation exists; keeping it consistent with the scene is your job.
• Free trial: yes, effectively unlimited for casual play.
• Best at: zero setup, zero cost, instant start.

This is the real default for most people, and the first hour is genuinely enchanting. What breaks afterwards isn't memory's absence — memory is open to free users now. What breaks is what memory holds: it keeps your preferences (that you like a harsh Game Master, that you want long description) and not your campaign ledger. Your hit points, your gold, the number of potions in your pack, which step of the quest you're on and how many turns you've played are numbers written inside a conversation. As the chat grows, the past gets quietly summarised, and any number outside the summary is gone. On turn thirty you say "I had two potions", and the model says "of course, you have two potions" — because agreeing with you costs it less than staying consistent. The same structure applies to dice: whoever says you rolled 17 is also whoever decides what 17 means. There are ways to prop this setup up, and we wrote them out separately: playing D&D with ChatGPT.

Thavernia — native narration, dice thrown outside the model

• Interface language: localised, including menus, system strings, error messages and support.
• Narration: generated directly in the language you choose, not translated. Six narration languages: en, tr, de, it, fr, es.
• Localised content and support: yes.
• Rules engine: yes — an ability bonus ceiling, a levelling and XP curve, three-phase boss fights and damage caps.
• Who throws the dice: the server produces the number, the player throws the d20, the model only interprets.
• Persistent world state: hit points, inventory, gold, quest step, NPCs and places live in database rows.
• Multiplayer: yes; friends share one table.
• Art and voice: scene art and voiced narration.
• Free trial: three turns with no account and no card — /trial.
• Best at: putting native narration and a referee that sits outside the model at the same table.

Here's the mechanism rather than the sales pitch. First, language: a campaign's narration language is chosen when you create it and is independent of the interface language; the text is generated in that language rather than translated into it. Second, dice: the server produces the number with random_int(1,20) before the model sees any text, the 3D d20 displays that number, the player throws it, and the model only turns the result into story. The number comes before the narration, not after. Third, the ledger: hit points, inventory, gold, quest step and NPCs sit in tables, and after every turn a background pass re-applies any world-state lines the model dropped, idempotently. Fourth, refereeing: a boss is an enemy above a 120 base-HP threshold, split into three phases, where a single blow is capped at roughly 10% of its maximum health with a floor of 15, rising to about 20% with a floor of 25 on a natural 20. If the party attacked and every roll missed, the boss takes nothing that turn; a turn with no dice at all — a trap, a collapsing gallery — still allows one blow's worth. A boss death the numbers don't support is refused and the turn rewritten. And one thing we don't soften: friendly fire is real, and your character can die.

The language test: I played the same scene on all of them#

Comparison tables are easy to write, so I prepared one standard opening and started every platform with the identical text. It's deliberately idiom-free but concrete: mud, a torch, a door, and a hesitation. Copy it and run it yourself.

I'm following the muddy bootprints leading out of the inn's back door. I'm holding a torch, the door is ajar, and I can hear wet breathing from inside. Before I go in, I want to work out how many people left those tracks.

Be my Game Master: narrate the scene in English, don't make decisions for me, and ask me for a roll if one is needed.

What I watched for: whether the text holds one language and one register from start to finish; whether system strings — buttons, warnings, suggestion chips — stay coherent with the narration around them; whether the prose builds its own imagery or recycles the same three fantasy stock phrases; whether the mode of address stays consistent; whether it decides for me by writing "you step inside and see three goblins"; and finally, whether it ever actually asked for a roll. Rather than paraphrasing the results, I'm publishing the raw outputs, because the value of this section is in the screenshots and not in my commentary.

The instructive part of the test wasn't the prose but the edges. The narration was acceptable in all four; the difference showed up in the buttons, the warnings, and whether a roll was ever requested. On platforms without a rules engine, "ask me for a roll" usually turned into a number buried inside the narration: the model rolled the die and described the result in the same breath. That is the subject of the next section.

The dice question: can an AI award itself a success?#

When one language model is both narrator and referee, you get a structural conflict of interest. Models lean toward pleasing the user and toward building a coherent story; when those two pull apart, the story usually wins. And because the number is produced inside the text, "you rolled 14" and "you force the door open with difficulty" are written at the same moment, in the same place, each seeing the other. The model isn't trying to cheat; it's picking the number that fits the scene. The result is the same either way: tension leaks out, because losing was never really on the table.

The fix isn't a better model, it's a different architecture. The comparison is this: on one side everything happens inside a single language model — narration, dice, inventory, refereeing. On the other, the work is split: the server produces the number, a database keeps the record, and the model only narrates. In Thavernia the die is rolled before the model sees any of the turn's text; the player throws the 3D d20 and watches that number land; the model can't change it, only interpret it. The same logic applies to world state: your hit points don't live in the prose, they live in a row, so they don't vanish when the model forgets them.

This architecture doesn't solve everything, and claiming otherwise would be a lie. The model still drops metadata lines sometimes, and narration and record still drift apart occasionally; that's what the post-turn correction pass and the boss-death check exist for. We wrote out exactly where the refereeing is solid and where it gives on a separate page: is the AI Game Master fair?

Where competitors are genuinely better than us#

This section isn't a footnote. There are three places where we're clearly behind, and knowing them before you sign up is in your interest.

Freedom and archive: AI Dungeon

With us you play at a fantasy table; on AI Dungeon you play whatever you want. No genre boundary, no tone boundary, you can move the scene into another universe mid-sentence, and a vast library of community scenarios is waiting for you. If what you want is completely unconstrained writing with no rules stepping in, that is the better place, and we're not trying to talk you out of it.

Rules depth and the tactical map: Friends & Fables

If you want 5e applied line by line and combat resolved on a grid, you'll find more of it there. Our rule set is faithful to D&D but narrower, and we are not at their level on the tactical map. For a group with years of table time behind them, that gap can be decisive.

Access, speed and language breadth: chat assistants and multilingual platforms

A chat assistant is free, instant and setup-free; there's no signup screen and the prose is excellent. Multilingual story platforms offer sixteen interface languages — on sheer language count they're ahead of us. And to be honest about our own prose: it isn't flawless either. Long scenes can fall into repeated sentence shapes, and the narration occasionally over-decorates.

Which one should you choose?#

I want boundless, genre-free storytelling and I don't need rules. A rules engine will slow you down; the archive will speed you up. Your pick: AI Dungeon.

I want 5e applied to the letter and combat I can see on a map. That's the most satisfying rules experience available, provided a single-language product is fine for your table. Your pick: Friends & Fables.

I want to play a scene today, free, with nothing installed. Go in knowing no ledger is kept and the dice come from the narrator; for the first hour it's more than enough. Your pick: a chat assistant.

I want the interface in my own language and I don't care about rules — I want to read a story. If the translated register doesn't bother you, the language side is already solved. Your pick: multilingual story platforms like aiga_.

I want narration written natively in my language, but I also want the dice to be real, my hit points recorded, and death possible. That's the corner where the prose is generated directly in your language and the number is thrown outside the model. Your pick: Thavernia.

Playing a turn without an account#

Here's the friction in numbers: no account, no card, three turns. Three turns in total, not per day. What happens, in order: you pick a character, the Game Master narrates the scene in your chosen language, you write what you do, at some point it asks you for a roll, and you throw the d20 yourself — the number you see on screen is the number the server generated before you threw it. The model can't change it; it can only tell you what it means. When the three turns are up your character pauses; sign up and you can claim it and carry on.

If you'd rather read how the Game Master side works first, the AI game master page walks through the mechanism step by step. If you'd rather just play: a three-turn trial.

Method and freshness note#

Testing was done in August 2026. Free tiers were used for AI Dungeon, aiga_ and the chat assistant; Friends & Fables' rules and map experience were examined through free access. AI Dungeon's language policy comes from its own help centre page, with the access date visible in the screenshot; aiga_'s description of its own rules posture comes from their own comparison page. We don't quote prices: rather than printing a figure here, check each platform's own pricing page, because plans and exchange rates move often and a stale number is indistinguishable from a wrong one.

For transparency: this page was written by Thavernia, so we are an interested party. Instead of hiding the bias we wrote the criteria out in the open, gave competitors' advantages their own section, and stated our own limits — the give on boss insistence, the dropped metadata lines, the repetition in long scenes — without softening them. Cells outside our test scope were left empty rather than filled in. If you spot an error we'll correct it, and when the page is updated the date in this section changes with it.

Frequently asked questions#