MCP vs API: which to choose
MCP rides a subscription you already hold. A provider key buys a specific model and tighter spend. The numbers, and what only the built-in chat can do.
Guide
Choosing an AI model for Familiar.
Familiar runs on the model you connect, so the model sets the ceiling. A stronger one holds a multi-step tool chain together more reliably, and a round of combat is made of exactly that. Most models also expose a reasoning effort dial: spend more thinking on a tricky ruling, less when you want the reply back fast. This guide is how to pick both.
Which model should you pick?
The short version: connect a current flagship from Anthropic, OpenAI, or Google and you will not have to think about this again. A capable mid-tier model covers a tighter budget and carries most of a normal session. A strong local model runs offline if you want your world to stay on your machine. You bring your own AI, so all three are yours to take.
Familiar names the models it can reach in the picker itself, because hosted line-ups change every few months. What does not change is the shape of the choice.
The words are the model's own: whether the read-aloud lands, whether your innkeeper still sounds like himself in hour three, whether the recap is worth reading. That, and the judgement when two conditions collide, is what a stronger model buys.
| Tier | Best for | The catch |
|---|---|---|
| Flagship | Hard rulings, messy multi-creature fights, anything that turns on judgement | Priciest per token, and the slowest to answer |
| Mid-tier | Most of a normal night: narration, lookups, NPC talk, routine combat | Less headroom when a rule gets tangled |
| Fast | Quick lookups and recaps where you want the answer back now | Thinnest margin on a long chain of tool calls |
| Local | Offline play on your own hardware | Needs a tool-trained model of 14B or larger and a discrete GPU to feel like a conversation, and combat wants a much bigger model |
Where do you get one?
Three routes, and they hand you different amounts of choice.
One key for everything: OpenRouter's list holds only models that can call tools, ranked by what its users ran most this week, and you swap between them without signing up anywhere else. The shortest path from nothing to a first reply.
A direct provider key reaches that provider's own catalogue.
Over MCP, you do not pick a model in Familiar at all. Your client picks, and Familiar runs the tools it calls.
Whichever you take, none of it is a commitment. Your campaign lives in Foundry journals and Familiar's memory, not in the model, so swapping provider halfway through a campaign costs you nothing but the time to paste a key. Swapping models is quicker in the built-in chat: a picker sits in its message box, and the model you pick there answers your next message. Until you pick, the Model row sits on Provider default and names the model behind it: the provider's own pick, not one Familiar recommends.
Where the model shows up
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.
Opening a fight is a chain of tool calls. Roll initiative, set the order, then play the first monster turn: move it, force a save, apply a condition, update the canvas. Every step is a chance to slip, and the slips compound. A model that gets each step right nineteen times in twenty still comes through that six-step chain clean only about three times in four.
On D&D 5e (2024) Familiar takes some of that risk off the model: the damage, the conditions and the dice run in code, so the arithmetic holds once the right tool is called. What the model still owns in that chain is the order and the reading: which tool to reach for next, whether a rule applies at all, whether a ruling survives two stacked conditions. That is why the model shows up most in combat.
On other Foundry systems Familiar's 5e guardrails do not run, so more of the night rides on the model itself: the language, the lookups, and the calls you hand back to the table.
Speed is the other half of the trade, and pricing pages never mention it. A flagship thinking hard is dead air at a table with people sitting in it. Playing solo you will barely notice. With four players watching you wait, you will.
Combat AI
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 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, and the model spends the extra effort only where you point it.
Familiar ships this on High. That is already the right place for combat and hard rules, so if you never touch the setting, you are where the fiddly nights want you to be.
Which rungs you see depends on the model you connected; over MCP the dial sits in your client, not in Familiar. On a direct Anthropic key, Claude Opus runs from Off to Max; through OpenRouter it gets the levels OpenRouter lists. Most others stop at High. Off appears only where the model's own list offers it: some models have no Off at all, and some bottom out at Minimal instead. Familiar shows only the levels your model accepts, so a shorter dropdown is the model's limit rather than a missing feature.
Combat & tricky rules: High
Leave it here for the multi-step bookkeeping, the work that earns the extra thinking.
Roleplay & exploration: Medium
NPC banter, reading a journal, narrating a room. Enough deliberation to stay in character and on the page, without spending it on a conversation.
Quick lookups & recaps: Low, or Off where you have it
Fetch a stat block, restate a rule, summarise last session. Off is the fastest and cheapest where your model offers it; otherwise take the lowest rung it has.
Turning Thinking up improves the odds on a hard sequence; it does not guarantee the answer. Past a point the model over-thinks a simple ask and you pay for tokens you did not need. Give it the room the task needs and no more.
Is it the model, the dial, or your prep?
Before you go and buy a bigger model, work out what went wrong. Most of what looks like a weak model is not one, and a flagship will not fix a page you never wrote.
| Symptom | Likely cause | What to change |
|---|---|---|
| A condition gets dropped, or initiative lands out of order | The dial, then the tier | Raise Thinking first. If it still happens on High, move up a tier. |
| It loses the thread halfway through a big fight | The tier | A longer tool chain needs more headroom. This is the one a flagship fixes. |
| It forgot an NPC from two sessions ago | Not the model | Nothing was written down for it to read. Give that NPC a journal page. |
| It invented a beat you never wrote | Not the model | The page it needed did not exist. Fix the prep. |
| Replies are slow and the table is waiting | The dial is too high for the moment | Drop to Medium for the quiet stretch, and raise it when the fight starts. |
More in Choices around the connection
Running a local LLM
Run the model on your own PC, so your world never leaves the machine. One settings tab and one switch. Without a discrete GPU, expect minutes per turn.
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.