Transcribing your sessions
SoonSpeak instead of type. Familiar transcribes the session in real time and saves it to a searchable Foundry journal you can edit later.
Guide
People keep asking what my own game actually runs on. This is it: one PC with Foundry, my wife across the table with her character sheet open on her phone, a microphone between us, and Claude, on the plan I already pay for, playing two of the three heroes and voicing every NPC out loud. Nothing here is a lab demo. It is the rig we play on, including the parts that took broken sessions to get right.
Three of us play. I sit where the DM always sits: I frame the scenes, make the rulings, and decide what the world does next. My wife plays Naia, her sorcerer, and nobody else touches that character. The AI plays the other two heroes, a barbarian and a bard, and has done for months. It also runs the monsters, voices the cast, and keeps the books.
The game is Ember, Foundry's official sandbox campaign, imported and prepped before anyone sits down. That choice matters more than any tool below it. Given nothing, a model invents, and inventions drift. Given a prepped campaign, it reads and runs what is on the page, and the world stays solid for months at a stretch.
Every piece is ordinary on its own. The work happens in how they connect: Foundry holds the world, Familiar gives the assistant hands inside it, and the assistant I already subscribe to does the thinking. There is no separate AI bill at my table. The only extra keys are for speech: one that gives the cast voices, one that turns the microphone into text.
| Piece | What it does |
|---|---|
| Foundry VTT | Self-hosted on the PC at the table. Holds the world: maps, sheets, journals, and dice everyone can see. |
| Familiar | The module inside that world, plus its MCP server. Every roll, token move, journal read, and voice line the AI produces goes through it. |
| Claude | Connected over MCP on my existing subscription, so there is no per-token bill. Any MCP assistant works; this is simply the one I run. |
| Character voices | An ElevenLabs key in Familiar's voice settings. The narrator keeps one fixed voice, recurring NPCs keep theirs, and lines play in the browser as the scene happens. |
| My wife's sheet | Her phone on the house Wi-Fi, logged into Foundry with the Sheet Only module, so her whole screen is just Naia. She rolls her own dice. Always. |
| The microphone | A desk mic and a Gladia key turn what we say into a running transcript the AI reads, so I talk instead of type. Familiar also transcribes inside its own chat window; my table plays over MCP, so a small script pipes the mic to the assistant instead. |
The rhythm took a few sessions to find. What survived is a short pre-flight, one hard rule about text and audio, and a wrap-up that makes the next session cheap to start.
Open the world, start the assistant, and run one spoken test line. Two minutes, and it catches a dead mic or a silent voice before the cold open instead of during it.
I frame the scene and rule the edge cases. The AI narrates, plays its two heroes, and voices the NPCs. Every spoken line lands in the Foundry chat first, word for word, so the written record always matches what was heard.
Initiative, monster turns, and the AI's heroes run through Familiar on real dice. My wife declares her own actions and rolls her own; the AI never decides anything for her character.
The AI writes the session log and one pickup paragraph: where the party stands, what is open, who is hurt. The next session starts by reading that back instead of by remembering.
Combat & AIVoice & Image GenerationKnowledge & MemoryNone of the discipline above comes from the model's good nature. It comes from standing instructions the assistant reads before every session, split into small files with one job each. The split is the point: a model follows a few focused rules far better than one giant document, and small files let me fix one habit without retyping the rest.
The portable half is public. The rules pack is the table-manners layer, and it works pasted into any assistant. On top of it sit files only my table needs: how each character speaks, the combat house rules, what is canon in the world, and which secrets never appear on any page a player can see. Where they live depends on your assistant: a Claude Project, a Gem, or a rules folder for the coding-style tools.
my-game/
rules/
table-manners.md the pack from /guides/dm-rules-pack
roleplay.md who speaks how, spoiler discipline
combat.md house rules, when to ask the table
world.md what is canon, what is off-limits
session-runbook.md pre-flight, the session loop, wrap-up
logs/ one file per session, newest is the pickup
The echo was the worst one. A far-field mic hears everything, including the AI's own voice lines coming out of the speakers, so the transcript filled up with its lines attributed to me, and the speaker labels collapsed into one voice. The fix was embarrassingly small: send the AI's audio to headphones, or pause the mic during playback. If your transcript ever starts talking to itself, that is why.
Mid-session debugging nearly killed the pacing twice. When a tool call fails, the pull to fix it right there is strong, and it is wrong: two attempts, then narrate the result by hand and repair it after the session. The table came for the story. Nobody came to watch me read error messages.
And a frozen fight does not always thaw the way you left it. We pause combats between evenings, and the world can shift in the meantime. So the wrap-up records the exact combat state, and resuming starts with checking the tracker against the log before anyone rolls. Thirty seconds, and more than once it has caught a tracker that had drifted.
You can assemble this in an afternoon, and the order matters less than people fear. Install Familiar in Foundry, connect the assistant you already pay for, and paste in the rules pack. That is a playable table. Voices, the microphone, and the phone sheet are upgrades you add the week you want them, not prerequisites.
The one thing no tool installs is the campaign. Bring a published adventure you love, prep it into journals the AI can read, and it will run the long game with you at whatever size table your house happens to hold.
Good AI D&D is good prep. Structure a published adventure, then hand the running to the AI.
You know these modules cold and never get to play one. Bring a published adventure, take a single character, and Familiar runs the DM's side around you: the monsters, the NPC voices, the 2024 5e rules.
Lay a published adventure into Foundry as journals, sheets, and a one-page outline, so the AI has the pages to run from.
Give an NPC who they are, what they know, what they want, and how they speak. Anchored to that, the AI voices them and fills the small edges itself.
A long campaign overflows any context window. Keep a searchable record in Foundry so the AI reads from your notes, not from a fading window.
Let the AI speak. Set a narrator voice and character voices, bring your own provider key, and Familiar voices each line in your browser as the scene plays.
Run combat with an AI that cannot cheat. Foundry's own dice enforce range, damage, and conditions, while you stay the DM.
Speak instead of type. Familiar transcribes the session in real time and saves it to a searchable Foundry journal you can edit later.
New to Familiar? I'm Ryan, the person who built it. The Discord is small and brand new, so if you join now I'll help you get set up myself.