Choosing an AI model for Familiar.

Pick a model

Think of the model as the brain that executes your prep. The journals, statblocks, and encounters are already on the page; the model is what reads them and works the tools to make them happen at the table. Prep fixes what runs. The model sets how well it runs.

One turn of combat is a multi-step tool sequence: roll initiative, set the order, apply a condition, force a save, update the canvas, then play the monster turn. Bigger flagship models follow that chain most reliably. A weaker one fumbles more of the steps, and the more conditions you stack, the more it slips. So the model shows up most where it counts, in combat and the fiddly bookkeeping you wanted off your plate.

You bring your own AI, so the choice is yours, and Familiar works with any model these providers expose. Want the most reliable table? Reach for the flagship tier. A capable budget model covers tighter spend, and a strong local model runs fully offline if you want nothing to leave your machine.

• Best reliability: Claude Opus 4.8 or Sonnet 4.7 (Anthropic), GPT-5.5 (OpenAI), Gemini 3.1 Pro (Google), Grok 4.3 (xAI), Mistral Large 3 (Mistral), or DeepSeek V3.2 (DeepSeek).
• Budget: DeepSeek V3.2 is capable and inexpensive.
• Offline: any strong local model through Ollama or LM Studio, nothing leaves your machine.

The flagship tier earns its keep in combat, where one turn is a chain of tool calls. A smaller model is fine for quiet scenes and lookups, and you can switch per session. Match the model to how mechanical your night will be.

Give it room to think

Reasoning effort, what Familiar calls Thinking, is how hard the model deliberates before it acts. Turn it down and the AI answers fast off the top of its head. Turn it up and it works the problem through first: reads the rule, checks the board, plans the sequence, then moves. You set the dial. The model spends the extra effort only where you point it.

It matters most when the AI has to chain several tools together. Combat is the clearest case, the same tool chain from above. Run it on low reasoning and it fumbles: initiative lands out of order, a condition gets dropped, a save resolves against the wrong DC. Run it on High and the model thinks through the whole sequence before touching anything, so the steps come out clean.

More is not always better. Reasoning effort is a lever, not a correctness guarantee. Past a point the model over-thinks a simple ask and you pay for tokens you did not need: more thinking means more tokens, and a bring-your-own-key session runs you roughly $0.50 to $3. So the goal is not to max everything. It is to give the AI enough room for the task in front of it, and no more.

1. 1

   Roleplay & exploration: Medium

   The balanced default for most gameplay. NPC banter, reading a journal, narrating a room. Enough deliberation to stay in character and on the page, without burning tokens on a conversation.

2. 2

   Combat & tricky rules: High

   Thorough reasoning, the model default, and the best setting for hard rules and tactics. This is where the multi-step bookkeeping lives, so this is where the extra thinking earns its cost.

3. 3

   Quick lookups & recaps: Low or Off

   Off is the fastest and cheapest. Fetch a stat block, restate a rule, summarise last session. One-shot answers that need no deliberation, so do not pay for any.

Thinking is a dial you control, not a promise we make. Spend it where the work is hard and save it where the work is simple. Familiar will not pretend Extra High turns a weak setup into a perfect game; it gives you the lever and tells you where it pays off.

How you connect changes who controls thinking

There is one wrinkle worth knowing: the reasoning dial does not always sit in the same place. It depends on how you connected Familiar to your AI. The two routes hand the controls to different places, and that catches people out.

On the built-in chat route, you bring an API key and pick everything inside Familiar. You choose the model in the settings, and a Thinking dropdown next to it sets how hard that model reasons, from Off through Low, Medium, and High up to the deepest tier. Both knobs live in one place, in front of you, and you change them whenever the night calls for it.

Over MCP, the controls move. Your AI client runs the model: Claude Desktop, ChatGPT, or Codex decides which model answers and how much it reasons before it does. Familiar just runs the tools that client calls in your Foundry world. So in MCP mode the reasoning dial lives in your client, not in Familiar, and you set it the way you already set thinking or reasoning effort in that app.

Neither route is better here. It is about where the lever sits. If you have not picked a route yet, the "Connect via MCP" and MCP vs API guides walk through both in depth, including who you pay and what each one costs.

• Built-in chat (API key): you pick the model and the Thinking level inside Familiar.
• MCP: your AI client picks the model and controls how hard it thinks; Familiar runs the tools it calls.

A simple recipe

If you would rather not tune this, here is a starting point that works for most tables. Run it for a session or two, then adjust to how your game actually feels.

Built-in chat exposes these levels directly in the Thinking dropdown. On MCP, read each step as the equivalent reasoning setting in your client.

1. 1

   Start on a flagship model at Medium thinking

   A strong current model with Medium reasoning handles the bulk of a session well: narration, lookups, NPC chatter, routine combat. It is the balanced default.

2. 2

   Bump to High when it gets hard

   A messy multi-creature fight, a stacked-conditions ruling, a tangled scene. Raise the dial for the moment, then ease it back once the table is past it.

3. 3

   Drop to Low or Off for quick stuff

   Short lookups, a quick recap, a name for a passing NPC. Light reasoning answers these fast and spends fewer tokens, which matters most on a metered API key.

On a tighter budget, swap the flagship for a cheaper capable model and keep it at Medium. Lean on it for most of the night, and only reach higher when a moment earns it.