Here is the dirty secret of running a virtual tabletop. Most of the job is plumbing.
Mid-combat, a token won’t move. An NPC’s attack bonus is off by two and I don’t know why. A status effect I swear I cleared last session is back, smug and persistent. Somebody needs a rules lookup on grapple (it’s always grapple). And while I patch that, I’m supposed to be prepping the ambush three rooms ahead and tracking whose turn it is.
I have downloaded every quality-of-life module the community has ever shipped. Dozens of them. They help. They do not save me.
Because here’s the math nobody tells you. The more time a GM spends on the Back End, the less they have for the Fun End: the part where you help five people tell their story. Every minute I spend fighting the software is a minute I’m not being a Game Master. I’m unpaid IT support for a game I’m also supposed to be running.
So I tried something stupid. I opened Claude for Chrome, pointed it at Foundry, and taught it to play Pathfinder. Actual Pathfinder. Rules, tokens, the works.
It learned. Slowly, clunkily, like a new player who’s read the book once. But it learned.
There was a ceiling. Claude in Chrome does one thing at a time. Click a token. Click again. Type. It’s bound to the same input I am, a mouse and a keyboard, one action per round-trip. No batching, no holding a six-step plan in its head while it works. Tell it to move the orc, swing the axe, apply the damage, and advance initiative, and you watch it do each step like a person, one careful click at a time. Useful. Not fast.
That’s when I got greedy. (Develops nefariously, as the man says.)
If Chrome was the mouse, Claude Code could be the hands. So I sat down and waged a one-man war on the English language: a design document that grew teeth and a name. The Automated Assistant Game Master. AAGM, because of course there’s an acronym.
The intent is easy to say and a nightmare to build. A Claude that lives inside Foundry. One you just talk to.
“Claude, move Orc 3 to Bob and hit him with the greataxe.” “Claude, Gem’s sheet corrupted again, fix her please.” “Claude, find the Sword of Epic Swordness and award it to John.”
And it does. That’s the time saver. The life saver is quieter. It’s the corrupted-actor fix at 9:58 PM with four people waiting, the thing that used to end a session early. Now it’s a sentence.
Now. You don’t hand an LLM the keys to your game world without thinking hard about what happens when it’s wrong. Because it will be wrong. The only question that matters is whether wrong is recoverable.
So AAGM has three hard rails, and they are load-bearing.
One. Claude cannot kill a player character. The damage path has an absolute floor of one hit point. It can drop you to your knees. It cannot put you in the ground. Lethal is a human call, mine, every time.
Two. Claude cannot destroy items. When it deletes something, the item doesn’t die. It ships to a box on a blank scene, sitting there fully recoverable, a lost-and-found nobody visits.
Three. Claude cannot truly slay a foe. Kill the goblin and a one-HP copy of it quietly appears on that same blank scene. Nothing is ever actually gone. The undo button is a place.
(There’s an unofficial fourth rail: talking Claude out of answering every problem with Fireball. Corrupted actor? Fireball. Incorporeal ghost? Fireball. Merge conflict? Fireball. The spell list is wider than it looks.)
Are those rails ironclad? No. I’d be lying, and you’d catch me. They’re good. They’re tested. They are not a mathematical proof, and anyone who tells you their LLM safety is a proof is selling something.
Which leaves one honest place to land this. Handing Claude Code the controls to your Foundry world could never, ever lead to a Machine Uprising inspired total party wipe.
Never.
Want to run a GM who doesn’t sleep? The module, relay, and docs are all up here: github.com/DatJavaClass/AutomatedAssistantGM. Clone it, bring Claude to your next session, and tell me what it breaks.
Just keep a human near the kill switch. You know. For story reasons.
