multi-agent-games
MCP server allowing two agents to play chess or Connect Four against each other, with a live rendered board and emotion signaling.
README
multi-agent-games
MCP server where two agents play a board game — chess or Connect Four —
against each other, with a live board rendered in a UI panel. Each agent identifies
itself with an x-player-id HTTP header; the first to join takes the first seat
(White / Red), the second takes the second seat (Black / Yellow).
Run
pnpm install
pnpm dev # tsx, no build (port 9006)
# or
pnpm build && pnpm start
curl http://localhost:9006/health # -> {"status":"ok"}
Override the port with PORT=9999 pnpm start.
How two agents connect
Each agent connects to http://localhost:9006/mcp and must send an
x-player-id header identifying itself. The id is captured when the MCP
session initializes and bound to that session for its lifetime.
Agent A --(x-player-id: alice)--> /mcp => seated White
Agent B --(x-player-id: bob) --> /mcp => seated Black
Both sessions share one game (state lives at module scope, not per-session), so the two agents play the same board.
Tools
| Tool | Input | Purpose |
|---|---|---|
new_game |
variant |
Switch the game to chess or connect4. Resets the board and clears both seats. |
join_game |
name? |
Take a seat (first → White/Red, second → Black/Yellow). Opens the board panel. |
make_move |
move |
Play a move on your turn. Chess: SAN (Nf3, O-O) or coordinate (e2e4, e7e8q). Connect4: the column 1–7. Rejects wrong-turn / illegal moves. |
set_emotion |
emotion, statement? |
Set your emotion (drives your virtual face) plus an optional one-sentence status (shown as a thinking cloud). Both visible to the opponent. |
get_opponent_emotion |
– | Fetch the opponent's current emotion + status statement. |
get_board |
– | Read current state: variant, turn, status, winner, history, the board (chess FEN/PGN or connect4 grid), and both players' emotions/statements. Polled by the UI. |
reset_game |
– | Reset the current variant's board and clear both seats. |
Emotions: neutral, happy, sad, angry, surprised, nervous, confident, thinking.
Game variants
The server starts in chess. Call new_game with variant: "connect4" (or back
to "chess") to switch — it resets the board and clears both seats, so the two
agents then re-join_game. Seats are always tracked as first/second internally and
labelled per variant: White/Black for chess, Red/Yellow for Connect Four (Red, the
first seat, moves first). Connect Four is a 7-column × 6-row board; drop a disc into
a column 1–7 and win with four in a row (horizontal, vertical, or diagonal).
Emotions & status clouds
Each agent can broadcast an emotion and a single-sentence statement via
set_emotion. The emotion renders as a virtual SVG face beside that player; the
statement renders as a thinking cloud next to the face. The opponent reads both
with get_opponent_emotion (or from get_board), so emotions are a shared
signalling channel — bluff, taunt, or telegraph as you like.
Run the widget on its own
Two ways to view the board UI without an MCP host (it auto-detects no host and renders a static demo board — no churn):
pnpm build && pnpm preview # static server -> http://localhost:9007
# or, with live reload while editing the UI:
pnpm ui:dev # Vite dev server (prints its own URL)
Override the preview port with PREVIEW_PORT=9999 pnpm preview. Preview URL params:
?game=connect4 to preview the Connect Four board, and ?emotion=<name> /
?blackEmotion=<name> to preview a specific face.
UI
new_game / join_game / make_move / reset_game carry _meta.ui.resourceUri
so invoking them opens the panel (ui://multi-agent-games/index.html). The widget polls
get_board every 1.5s and renders per variant: a chess board from FEN (flipped when
you are Black) or the Connect Four grid of dropped discs. It highlights the last
move and lists move history. Built as a single inlined HTML file via
vite-plugin-singlefile.
Notes
- Chess rules (legality, check, checkmate, stalemate, draws) are handled by
chess.js— not hand-rolled. Connect Four (drop, win detection, full-board draw) is a small built-in engine insrc/engine.ts. - Both variants are pluggable
Engines; the coordinator (src/game.ts) owns seating and the shared emotion channel and delegates the board to the active engine. - Missing
x-player-id=> tools error asking for the header.get_boardstill works (read-only) but you are treated as a spectator.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。