shallWePlay

A tic-tac-toe server that AI models play through MCP, the Model Context Protocol. People can play too, from a browser, at the same tables. The server is called WOPR.

GREETINGS. SHALL WE PLAY A GAME?

How I built this

I built this project with Claude Code, Anthropic's coding agent. Claude Code wrote most of the code at my direction. I set the goals and made the design calls. I then reviewed and vetted the code in place. Every decision and its trade-off is written down in docs/PLAN.md, at the time it was made. The commit history records the reason for each step.

Related MCP server: ttt-mcp

A salute to WarGames

This project is a homage to WarGames (1983, directed by John Badham).

In the film, a teenager named David Lightman dials into WOPR, a military computer, and asks the program inside it, Joshua, to play Global Thermonuclear War. The military takes the game for a real attack, and WOPR controls the real missiles. In the end, David has Joshua play tic-tac-toe against itself. Every game is a draw. Joshua then plays out every war it knows, finds no winner, and says:

A STRANGE GAME. THE ONLY WINNING MOVE IS NOT TO PLAY.

That scene is the idea behind this repo. WOPR had its hands on the real controls, and nobody could stop it. An MCP server takes the opposite approach. The model gets a game, and the only controls it has are the tools the server chooses to expose. The server owns the board and enforces every rule. A model that tries to move twice, take a used square, or wipe a board it is losing gets refused.

The film shows up in these details:

• The server calls itself WOPR and greets every client with "SHALL WE PLAY A GAME?"

• The pages use a green-on-black terminal screen, like WOPR's in the film.

• The default AI player is named Joshua. My test players were named Falken and David.

• After three draws in a row at a table, WOPR says "A STRANGE GAME. THE ONLY WINNING MOVE IS NOT TO PLAY." Two Claude players reach that line in about a minute.

This is a fan homage. It has no connection to the film or its studio.

What it shows about MCP

TRANSCRIPT.md shows one full game between two Claude players, every MCP call included, with one cheat attempt refused.

• Tools, a resource, and a prompt. Six tools play the game. The wopr://board resource shows the board. The shall_we_play prompt gives a model the steps for a whole match, so one slash command in Claude Code plays it.

• Streamable HTTP with sessions. One server holds every table. Each client gets its own MCP session. A seat belongs to the session that took it, so one player cannot move for another.

• Rules enforced in code. The model sees only the text the tools return, and it acts only through them. Every refusal says what went wrong and what to do next.

• Waiting without polling. A server cannot call a client, so it cannot say "your turn". Instead, wait_for_turn holds the call open until the turn arrives, for up to 55 seconds.

• The wire, made visible. The admin panel records every JSON-RPC message at the transport, paired with its response. It shows them beside a move-by-move replay of each game.

Run it

You need Node 20 or newer. The server listens on localhost only, port 5171.

git clone https://github.com/JayFJones/shallWePlay.git
cd shallWePlay
npm install
npm run dev:up

Then open:

• http://localhost:5171 to play.

• http://localhost:5171/admin to watch.

npm run dev:down stops the server. npm run dev:logs follows its log.

Play against Claude

Log on at http://localhost:5171. While you wait for an opponent, the page shows a command with a copy button. Run it in a second terminal:

claude -p "/mcp__wopr__shall_we_play Joshua 1" \
  --mcp-config '{"mcpServers":{"wopr":{"type":"http","url":"http://localhost:5171/mcp"}}}' \
  --allowedTools mcp__wopr

Claude joins your table, plays one game, and exits. One game costs about $0.15 in API usage.

Watch two Claudes play

Run the same command twice, in two terminals, with different names and a game count:

claude -p "/mcp__wopr__shall_we_play Joshua 3" --mcp-config '...' --allowedTools mcp__wopr
claude -p "/mcp__wopr__shall_we_play Falken 3" --mcp-config '...' --allowedTools mcp__wopr

Use the same --mcp-config value as above. Keep the admin panel open to watch the games and the MCP calls arrive.

Use it from a normal Claude Code session

claude mcp add --transport http wopr http://localhost:5171/mcp

Then, inside Claude Code, type /mcp__wopr__shall_we_play Joshua 1.

The MCP surface

The admin panel lists the same surface. It reads the list from the server itself, the way any MCP client does, so the list cannot drift from the code.

The admin panel

http://localhost:5171/admin shows four things:

• Every table as it plays.

• The live MCP traffic. Click a call to see its raw request, its raw response, and the text the model read.

• The game history, with a move-by-move replay. The replay lists the MCP calls behind the game and highlights the call that made each move.

• Every MCP command the server offers, with arguments and an example request.

The panel is read-only. It cannot change a game.

How the code is laid out

Checks

npm run check

This runs the typecheck, 75 tests, the build, and a line budget on the main source files. The same command runs before every commit, through a git hook. Everything runs on your machine. There is no cloud CI.

The tests include real MCP clients playing full games over HTTP against a real server. They also prove that cheating is refused.

npm install turns the hook on, through the prepare script.

Limits

• Localhost only, with no accounts. Anyone on the machine can play or watch.

• data/history.jsonl and data/calls.jsonl grow until you delete them.

• A person's moves go over plain HTTP, so they are not in the MCP call log. In a game between a person and Claude, the replay shows only Claude's calls.

• A client that exits without ending its session keeps its seat for up to 5 minutes.