MCP vs API: which to choose.

The two ways to connect

Two routes, and the difference is who you pay and how. The first connects Familiar to a chat client you already subscribe to, Claude, ChatGPT, or Codex, through the Model Context Protocol. Your subscription is already paying the AI bill, so the connector adds no separate per-token charge. The "Connect via MCP" guide covers the Setup Wizard that generates the config for you.

The second is the built-in chat inside Foundry, most often run with a provider API key: pick a provider, paste a key, and start talking to your game. The key goes straight from your browser to the provider, which bills you per token for what you use. No subscription, no app in the middle, just your key and the model you chose. The same chat can also point at a local model like Ollama or LM Studio, which runs free on your own machine.

Both routes reach every tool. So the choice is not about features. It is about cost, and which one fits how you play.

MCP uses a subscription you already pay for. The API key bills you per token. Same tools either way; the question is which payment model is cheaper for your table.

The cost reality

Start with what is fixed. Familiar itself is a flat fee, the same whichever route you pick. On top of that sits the AI cost, and that is where the two routes split.

MCP runs on a subscription you already hold. A Claude or ChatGPT plan is a flat monthly price, and connecting it through MCP does not add a per-token bill on top. Anthropic connectors work on the Pro, Max and Team plans; ChatGPT and Codex offer the MCP connector on Plus, Pro and Team. For sustained roleplay, a session a week over months, a flat subscription usually buys more play per month than metered tokens do.

A paid provider key is the other model. There is no free tier on the hosted providers: each one bills you per token, in and out, for every message. Rates are directional and they change, so check current provider pricing before you commit, but as a rough shape a Sonnet-class model runs near $3 per million input tokens and $15 output, an Opus-class model near $5 in and $25 out. A single session is small money. Volume is where it adds up. The exception is a local model in that same built-in chat, Ollama or LM Studio, which costs nothing per token but runs on your own hardware.

One distinction that trips people up: a subscription and the API are billed independently. A ChatGPT Plus plan does not hand you API credits, and an API balance does not unlock the chat app. You are choosing one tap or the other, not topping up a shared pool.

The subscription route is not unlimited. It uses your plan's normal message and usage quota, so a marathon session can still hit your plan's rate limits. It is no extra per-token bill, never free usage.

When to choose MCP

MCP is the easy call when the subscription is already a sunk cost. You pay for Claude or ChatGPT anyway, so pointing it at your Foundry game is found value, not a new line on the bill.

• You already pay for a Claude or ChatGPT plan. The AI cost is covered; the connector just puts it to work at the table.
• You play regularly. A flat monthly price suits a weekly game better than a meter that climbs with every session.
• You want a predictable cost. One subscription, one known number each month, with no per-token surprises after a long night.
• You are happy inside your plan's usage limits. A typical session fits; just know that a very heavy one can brush the plan's rate cap.

When to choose the API key

The API key wins when you want control, or when you simply do not hold a chat subscription to lean on. It is the simplest path to start. It also runs the built-in chat inside Foundry, and a handful of conveniences live there only: live transcription, auto-pilot for NPC turns, automatic memory, slash commands, and drag-and-drop attachments. Voice, scenes, and image generation are not on that list; those work over MCP too, with the media rendering in your Foundry tab and generated images coming back inline.

• You want a specific model your subscription does not include. A key reaches the full provider catalogue, not just what the chat plan ships.
• You want fine-grained control over the model and the spend, paying only for the tokens a session actually uses.
• You have no chat subscription. With light or occasional play, a per-token bill can come in under a monthly plan you would barely use.
• You run a headless or automated setup where a desktop chat app is not in the picture.

Pick the cheaper tap and start

If you already pay for Claude or ChatGPT and play most weeks, connect through MCP and let the subscription you hold do the work. If you want a particular model, tighter control, or you have no chat plan to lean on, paste an API key and pay for what you use. Neither is locked in; you can switch routes later without redoing your prep.

For the step-by-step on the subscription route, see the "Connect via MCP" guide. To get going right now, install Familiar in Foundry and connect the AI you already use. If you are weighing the two against your own playing habits, the Discord is a good place to ask.