Tachikoma Game Design MCP

Local and read-only MCP server for Claude to query a Chroma library about narrative and video game design. This repository contains only the public runtime; extraction, chunks, and embedding generation live in tachikoma-rag-infra.

Features

• Multilingual semantic search with BAAI/bge-m3.

• Results with book, page, text, and similarity.

• MCP transport via stdio, no open ports.

• Offline operation with a pre-downloaded model.

• Read-only tools and protection against embedded instructions.

• Configurable index path, no infrastructure coupling.

Related MCP server: Project RAG wiki

Contents

mcp_server.py                         Servidor MCP.
claude_desktop_config.example.json    Configuración de Claude Desktop.
requirements.txt                      Dependencias de ejecución.
pyrightconfig.json                    Configuración del análisis estático.
tests/test_mcp_server.py               Pruebas del runtime.

PDFs, texts, chunks, embeddings, and Chroma databases are not part of this public repository.

Installation

py -3.12 -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt

Providing an index

Tachikoma needs a Chroma collection created with the same embedding model. You can generate it via the private repository tachikoma-rag-infra or receive it through an authorized channel.

Available variables:

The model must be downloaded to the local cache before starting the MCP. The server uses local_files_only=True and will not download files during a conversation.

The log is rotating, does not store query text, and records model loading times, Chroma connection, embedding, vector search, and total execution.

Configuring Claude Desktop

1. Copy the entry from claude_desktop_config.example.json into mcpServers.

2. Replace the repository, Python, and Chroma paths.

3. Completely close Claude Desktop and reopen it.

On Windows, the configuration is usually at:

%APPDATA%\Claude\claude_desktop_config.json

Tools

• search_narrative_library: semantic search with optional filter by book.

• read_book_page: retrieves chunks of a page in order.

• list_library_books: lists indexed books.

• library_status: checks model, collection, and number of chunks.

Tests

python -m unittest discover -s tests -p "test_*.py"

Optional static analysis:

uv tool install basedpyright
basedpyright --pythonpath ".venv\Scripts\python.exe" -p pyrightconfig.json

Privacy

The index may contain document fragments. Do not publish Chroma or distribute indexes created from copyrighted material without authorization.