Claude Desktop and Claude Code connect to Foundry with no agent app in between.

Claude Desktop or Claude Code?

Both are MCP clients Anthropic builds itself, both run on the same Pro, Max or Team plan, and both drive the same Familiar server in the same world. The choice is mostly ergonomics: where you would rather type, and what you want to see come back.

One thing Desktop does that Code does not is start Familiar's server when the app launches, not when you open a chat. That is no reason to keep one of them closed: Familiar listens on one local port, and a second server started by the other client attaches to the first one and runs through it, so Desktop and Code can sit open side by side against the same world.

For Claude Code the whole setup is the line below, at user scope. Leave the placeholder secret alone: Familiar's wizard mints the real one and hands back the finished command.

claude mcp add familiar --scope user --env FAMILIAR_WS_SECRET=<your-secret> -- npx -y familiar-vtt

• Claude Desktop and Claude Code: the setup stepsWhich buttons in which client, and the fixes when a connection stalls.
• Install Familiar in FoundryThe module both clients connect to. Package browser or manifest URL.

What Claude reaches inside your world

With the connector live, Claude can read and act on what your world already holds. The journals you imported, the character sheets, the statblocks pulled from the module, the scene open on the canvas. Ask who the innkeeper was and it opens the journal entry rather than reaching for a plausible name; ask what the party has walked into and it reads the tokens actually on the map.

That reach is also the boundary. Claude answers from the pages your world already holds, which is what makes a name or a stat block worth trusting. It is also why the campaign has to arrive from somewhere else: a published adventure you imported and prepped. Claude runs those pages without rewriting them.

• Journals & Notes
• Characters & Items
• Scenes & Tokens

• Getting a published adventure into FoundryLaying a module into your world as journals, sheets and a one-page outline.
• The rules I give my AI DMRoll discipline, player agency, and what to do about a model that cannot roll: the rules to paste in before play.
• How to play D&D with AIThe method the prep and the rules slot into: what you bring, and what Claude runs.
• Running a different game system?The systems with a page of their own.

Check the link, then ask for something real

Whichever client you picked, the first line to send is the product's own, the one both wizard cards close on. If Claude answers with your world's name, the link is live, and everything under it is safe to try.

Before that line, check that the client is holding the connector. On Desktop, quit the app and open it again once the config is pasted. Under Settings, Developer, familiar then shows as running, and the + button under the chat box lists it under Connectors.

On Code nothing needs reloading, Foundry included. Familiar's MCP / Subs tab turns green on its own once a session has started the server, and /mcp inside that session shows the same status. The first start on either client downloads the server, so give it a minute.

After that, each ask below leans on something this page has already claimed: the sheets and statblocks in your world, the picture row of the table, the scene on the canvas. The portrait line needs an image provider in Familiar's settings, and on Desktop the bundle line further down as well.

• The verify line

  Use Familiar's get-world-info tool

  Copy The verify line
• Read a sheet

  Read <character>'s sheet and tell me their AC, their current hit points, and which spells they have prepared.

  Copy Read a sheet
• Pull a statblock from the module you imported

  Find <monster>'s statblock in the adventure I imported and tell me what it does on its turn.

  Copy Pull a statblock from the module you imported
• Fetch a picture

  Generate a portrait for <NPC name> and set it as their sheet image.

  Copy Fetch a picture
• Move the table to a scene

  Switch the table to <scene name> and tell me which tokens are standing on it.

  Copy Move the table to a scene

Where only Claude gets stuck

Desktop's faults come from reading things once, at launch. Code's come from the command line. Nothing on this list shows up on another client; the faults every client shares have one home, the connect guide, linked below.

• Settings, Developer shows no familiar after the paste

  Closing the window is not quitting. Desktop keeps running in the tray, so the entry you pasted has not been read yet. Right-click the tray icon and choose Quit on Windows, or press Cmd+Q on a Mac, then open it again.

• On Windows it is still missing after a proper restart

  Some Windows installs open one config file and read another. Look for a second claude_desktop_config.json under %LOCALAPPDATA%\Packages, in the folder there whose name starts with Claude_, at LocalCache\Roaming\Claude, and put the same entry in that one.

• Ask Desktop for a portrait, a voice or a battle map and nothing happens

  Desktop never looks at its tool list again mid-session, so it cannot pick up a tool it did not start with. The section below carries the one line that switches those three on.

• claude mcp add refuses the line: Invalid environment variable format: familiar

  The server name went in after the secret. That flag takes a list, so a name placed behind it is read as one more variable and rejected. Put familiar straight after add, the way the command higher up this page has it, or copy the finished line from the wizard card.

• claude mcp list says Failed to connect right after the add

  Adding the entry saves it and starts nothing. That check is what first runs the server, and while that first download is still going it reads as failed. Run it again a minute later.

• The failures every MCP client sharesSafari, an antivirus on Windows, a hosted game, and a server row reading Not running: answered once, for every client.

Bringing an Anthropic key instead

Some tables would rather not have a client open at all. Familiar takes an Anthropic key directly, and Claude then answers from inside the Foundry tab itself: no config file, no second process to keep alive. That key is metered by Anthropic, per token. Familiar itself is a subscription of its own, billed apart from anything Anthropic sends you.

The key buys one thing the plan route cannot. On a direct Anthropic key, Familiar renders the longest reasoning ladder it has anywhere. Those rungs are not inferred from a model name: Anthropic reports what each model accepts, and the dropdown is that answer. How far down the ladder goes depends on the family, since some Claude models let you switch thinking off entirely and some do not offer that rung at all. Route the same model through an aggregator and that ladder gets shorter.

• How hard to let Claude thinkWhat each rung is for, and where the dial sits on each route.
• Plan or key: which bill suits your tableThe per-session numbers, and when each route works out cheaper.