MCP for AI RPG: How to Run a Campaign on an Engine Outside the Model

An MCP server is a small program that keeps your game’s rules and records outside the chat window, and lets the AI you already use act as its narrator. The model writes the prose. A database somewhere else remembers that you sold the horse in session three, that the reeve’s daughter is still angry with you, and that it is now raining.

That is a genuinely different architecture from everything else in AI roleplay, and it is almost entirely undocumented for players. Search for it and you will find GitHub repositories written for developers, protocol specifications, and a scattering of tools for building games rather than playing them. Nobody has written the version for the person who just wants to know whether this fixes their campaign falling apart.

So here it is: what the protocol actually is, how it fits with the setup you probably already have, what exists that you can play today, and — the part the repositories will not tell you — what it costs.

MCP is an open standard originally published by Anthropic. Claude is a trademark of Anthropic, PBC; ChatGPT of OpenAI. Arcanum is not affiliated with either, and the products named below are referred to for identification only.

What MCP Actually Is

MCP stands for Model Context Protocol. Strip away the specification and it is a shared language that lets an AI assistant talk to an outside program: the program advertises a list of things it can do, and the model can choose to call them mid-conversation and read back the results.

It was built for connecting assistants to calendars, codebases and ticketing systems. The roleplay application is a side effect, and a lucky one, because the thing MCP is good at — letting a model act on a system it does not itself contain — happens to be the thing AI roleplay has been missing.

A useful comparison: a character card or a system prompt is text you hand the model, which it reads and interprets. An MCP server is a program the model can operate. The difference is authority. Text describes how combat should work and hopes the model applies it consistently. A server runs combat and hands back the result, and the model’s only job is to describe what happened.

In practice, a roleplay MCP server exposes actions along the lines of create character, move to location, roll a skill check, apply damage, advance the clock, save. The model calls them the way you would click them in a game, and narrates the outcome.

The Three Layers of an AI RPG Setup

If you have played inside ChatGPT or Claude before, you already own part of this stack. It helps to see all three layers separately, because they solve different problems and people routinely expect one to do another’s job.

Layer one — the rules. Your system: dice conventions, how combat resolves, what the GM is and is not allowed to do. This lives in a system prompt, or in a Claude Skill, which loads a ruleset on demand. We covered that layer in full in Claude Skills for tabletop RPG.

Layer two — the campaign container. The world, the cast, the history so far. A Claude Project or a Custom GPT holds this: persistent instructions plus a knowledge file, so you are not re-pasting your setting every session. The Claude RPG guide is the full treatment.

Layer three — the state. What is true right now. Your hit points, your coin, who owes you a favour, which door you left unlocked. This is the layer that has never had a home. In every prompt-based setup it lives in the conversation itself, which means it is competing for room with everything else you have said, and losing a little every turn.

MCP is a layer-three answer. It is not a replacement for the first two — you still want a ruleset, and you still want somewhere for your world document to live. It is the piece that was missing underneath them.

What Changes When the Engine Sits Outside the Model

We have written at length about why a prompt cannot hold a number reliably, and the short version is that a prompt is not a database — everything it establishes lives in the same finite, shifting window as everything else. That argument, with the scoring data behind it, is in why AI RPG mechanics break down, and it ends on a line worth repeating here: you cannot install a state layer into a product that lacks one.

MCP is the exception. It is the one architecture available to a player that actually installs one.

What that buys you is narrow and real. Facts the server owns do not decay, because they were never in the context window to begin with. Your inventory is correct on turn four hundred for the same reason it was correct on turn four: something outside the conversation is holding it, and the conversation’s length is irrelevant to it. Compare that with the ordinary failure curve described in why your AI campaign falls apart at turn 50, where the drift is not a bug anyone can patch but a property of the architecture.

What it does not buy you is prose. The model is still the model. Its voice still wanders on a long night, characters still soften toward agreeableness, and the narration still needs the same steering it always did. Moving state out fixes bookkeeping. It leaves the writing exactly where it was.

What You Can Actually Play Today

Honestly: not much. This is a real architecture with almost nothing built on it, which is why it is worth understanding now rather than in a year.

WyrdTale is the only MCP roleplay platform we have tested and scored3.0 out of 5 in our directory. It is a connector rather than a website: you install it into Claude or ChatGPT, and a server holds entities, saves, factions, combat, time and weather while the model you already pay for narrates. It reaches Memory — 4 and Longevity — 4, which is the joint-highest anything has scored on either axis across all 28 scored products, and it is the clearest evidence the architecture does what it claims. The full review covers where it falls down.

Beyond that, the field is open-source and early. role-playing-mcp-server is an MIT-licensed server exposing roughly seven tools for creating and progressing a game — genuinely an MCP roleplay server, and genuinely a small hobby project with no long-run testing behind it. We have not played it and are not scoring it; it is a pointer, not a recommendation.

One warning if you go looking. Several projects that dominate searches for MCP roleplay are not MCP servers at all. claude-rpg, for instance, is a well-made narrative RPG with its own simulation engine — but it is a command-line application you install with npm, not something you connect to a chat assistant. It may be exactly what you want; it just is not this. Check whether a project actually speaks the protocol before assuming it plugs into your Claude account.

How to Connect One

The mechanics are simpler than the surrounding jargon suggests. For a remote MCP server in Claude:

  1. Open settings and go to Customize → Connectors.
  2. Choose Add custom connector.
  3. Give it a name and paste the server’s remote MCP URL.
  4. Authorise it if the server asks you to, and start a chat.

That is the whole connection. From there the model can see the server’s actions and will start calling them when the fiction calls for it.

Two things make the difference between working and working well. First, pair the connector with a Project so your world document and standing instructions are loaded every session — the connector supplies the engine, the Project supplies the setting. Second, if the server ships a GM guide or recommended instructions, actually use them; they are what stops the model treating a perfectly good engine as decoration and improvising around it.

What It Costs You

The repositories will not tell you any of this, so:

Setup friction is real. This is the least beginner-friendly way to play an AI RPG that exists. You are installing a connector, possibly running software, and reading documentation written for developers. Compared with opening a character-chat site and typing, it is a different world.

Money is less of an issue than people assume. Custom connectors using remote MCP are available on every Claude plan, Free included — free accounts are limited to one custom connector. Projects are likewise available to all users, capped at five on Free, with Claude’s enhanced project knowledge reserved for paid tiers. So the entry ticket is smaller than it looks. The server itself may charge separately, and that is the cost to actually check.

Every action is a round trip. The model calls the server, waits, reads the result, then writes. That is slower and chattier than a model simply making something up, and on a wordy turn you will feel it.

And the big one: persistence does not buy you authorship. This is the finding that genuinely surprised us. The two platforms in our directory that reach the memory and longevity ceiling both score Player Agency — 1, arriving there by opposite routes — one by refusing to let you create freely, the other by writing your character’s actions for you. Our long-campaign rankings lay out both cases. A server holding authoritative state turns out to make it easier, not harder, for a game to decide what your character does next. If being the sole author of your own actions matters most to you, check that specifically before committing.

Who Should Skip This

Skip it if you are new to AI roleplay — start with a plain chat or a Project and come back when you know what you are missing. Skip it if your sessions are one-shots or short arcs, because state decay never gets a chance to bite and you would be paying setup cost for a problem you do not have. Skip it if the writing is the point: MCP does nothing for prose, and a well-engineered prompt in a good model will read better than a mediocre engine with a database.

Take it seriously if you run long, and if what breaks your campaigns is the game forgetting rather than the model writing badly. That is the exact failure this architecture removes, and nothing else available to a player removes it as cleanly.

Frequently Asked Questions

What is an MCP server in AI roleplay? MCP stands for Model Context Protocol, an open standard for connecting an AI chat assistant to an external program. In roleplay, an MCP server is a small program that holds the game — characters, inventory, factions, the clock, combat state — in a database outside the conversation, and exposes actions the AI can call. The model narrates; the server decides and remembers.

Do I need a paid plan to use an MCP RPG server? Not necessarily. Custom connectors using remote MCP are available on every Claude plan including Free, with free accounts limited to one custom connector. Projects are also available to all users, capped at five on Free. What a paid plan buys is headroom, not entry. The server itself may have its own pricing, which is separate from your AI subscription.

Is an MCP RPG better than a Custom GPT or a Claude Project? It is better at one specific thing: remembering. A Custom GPT or Project keeps your rules and world document persistent, but the running state of the game still lives in the conversation, so it degrades as the conversation grows. An MCP server keeps that state in a database instead. For a one-shot or a short arc the difference barely shows. For a campaign measured in months it is the whole ballgame.

Does MCP stop the AI forgetting my campaign? It stops the game forgetting. It does not stop the model drifting. Facts held by the server — who is alive, what you own, which faction hates you — survive indefinitely because they were never stored in the context window. Tone, voice and narrative texture still come from the model and still wander over a long session. Moving state out fixes the bookkeeping problem, not the prose problem.

What MCP roleplay servers can I actually use? The field is very small. WyrdTale is the only one we have played and scored, at 3.0 out of 5 in our directory. Beyond it there are early open-source servers on GitHub, most of them hobby projects with a handful of tools and no long-run testing. Be careful when searching, because several projects with RPG in the name are ordinary command-line games rather than MCP servers at all.

Does moving the engine outside the model give me more control over my character? No, and this is the finding that surprised us. The two platforms that reach the highest memory and longevity scores in our directory both score 1 out of 5 on Player Agency, and they get there by different routes. A server that holds authoritative state can just as easily narrate your character’s actions for you. Persistence and authorship are separate problems, and solving one does not solve the other.


MCP is the first architecture in AI roleplay that gives the game somewhere to live that is not the conversation, and the scores bear out that it works for exactly the problem it targets. It is also early, thin on the ground, and no help at all with the things a model is bad at. If you want to see how the measured field compares axis by axis, the benchmark publishes every score and the methodology explains how they are produced — and if you would rather have a finished, engineered game than assemble one, the Arcanum Originals are free.

Product names and trademarks are the property of their respective owners. Arcanum is independent and unpaid; no platform named here has paid for or influenced its coverage.