gamecode-rag
Enables semantic search and call-graph exploration over decompiled Unity Mono C# codebases, with hybrid retrieval, LLM re-ranking, and lazy-loaded project indexes.
README
gamecode-rag
MCP server for semantic search + call-graph over decompiled Unity Mono C# codebases.
Use it as the “library” next to the live bridge. Mono only — for IL2CPP use the decompiler, not this.
Related projects (same suite)
| Repo | Role | When |
|---|---|---|
| This — gamecode-rag | Semantic search + call graph over dumped Mono C# | Game is Mono |
| bepinex-mcp | Live Unity bridge (get/set/patch/watch) | Game is running with BepInEx |
| il2cpp-decompiler | Static IL2CPP decompile (needs Il2CppDumper) | Game is IL2CPP |
What it does
| Tool | Purpose |
|---|---|
list_available_projects |
List ingested game indexes |
code_search_and_rerank |
Hybrid search (vectors + symbols) → LLM re-rank top snippets |
code_graph_search |
Callers / callees for a method id |
ingest_new_project |
Build an index from decompiled sources or a Mono assembly path |
Queries for code_search_and_rerank should be full natural-language questions (not keyword bags). Symbol names (TakeDamage, PlayerController) still work well thanks to hybrid search.
How search works
- Hybrid retrieval — dense embeddings + keyword/symbol match over method/class ids, fused with RRF
- LLM re-rank — scores the broad set down to a short list
- Call graph — optional follow-up via
code_graph_search
Indexes under PROJECT_DATABASES/ are lazy-loaded: startup only scans folder names; a project’s vectors enter RAM on the first search/graph call for that project_id.
Embedding models
Default (OpenRouter): qwen/qwen3-embedding-8b — strong on code/retrieval, long context (~32k), and typically cheaper than OpenAI text-embedding-3-small on OpenRouter.
Local embeddings (optional): any OpenAI-compatible /v1/embeddings server:
EMBEDDING_BACKEND=openai_compatible
EMBEDDING_BASE_URL=http://127.0.0.1:11434/v1
EMBEDDING_MODEL=qwen3-embedding:0.6b
Works with Ollama, LM Studio, vLLM, TEI, etc. Re-ranker still uses OpenRouter by default (OPENROUTER_API_KEY + RE_RANKER_MODEL); if the key is missing, search skips re-rank and returns hybrid order.
Default re-ranker: deepseek/deepseek-v4-flash — typically cheaper and stronger at code relevance than openai/gpt-4o-mini. Override with RE_RANKER_MODEL if you prefer another OpenRouter chat model.
Use the same EMBEDDING_MODEL for ingest and query. Changing models requires re-ingesting every project (old indexes won’t load).
Requirements
- Python 3.10+
- .NET 9 SDK (for the Roslyn code-graph tool)
- Embeddings: OpenRouter or a local OpenAI-compatible embed server
- OpenRouter key recommended for LLM re-rank (optional if you accept hybrid-only ranking)
- Per-game index under
PROJECT_DATABASES/<project_id>/(you create these; not shipped)
Setup
cd gamecode-rag # this repo
python -m venv .venv
# Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
# edit .env → OPENROUTER_API_KEY=... (and optional local EMBEDDING_* — see above)
# Build the C# Roslyn parser (required for full ingest)
dotnet build tools/roslyn-parser/RoslynCodeGraph.csproj -c Release
Build an index
Expect this to take a while. Indexing a full game is not a quick script:
| Step | What happens | Rough time |
|---|---|---|
| Decompile (if needed) | ilspycmd dumps .cs from Assembly-CSharp |
Often minutes; large Unity games can be slow or need retries |
| Roslyn graph | Walks every file, builds methods + call edges → code_graph.json |
Often several minutes |
| Embed + ingest | OpenRouter embeddings for each method/chunk | Often many minutes to tens of minutes (size + API rate limits) |
You only do this once per game (or when you re-ingest). After that, search is fast. Leave the process running and don’t cancel mid-embed unless you mean to restart from that step.
Option A — folder of decompiled .cs files
# 1) Roslyn → code_graph.json (can take several minutes)
dotnet run --project tools/roslyn-parser -c Release -- \
--project-path "D:\path\to\decompiled\Assembly-CSharp" \
--output "PROJECT_DATABASES\my_game\code_graph.json"
# 2) Embed + call graph (usually the slowest step)
python ingest_code_graph.py --project-id my_game --source PROJECT_DATABASES/my_game/code_graph.json
Option B — MCP one-shot (ingest_new_project with assembly_path or source_code_path): decompiles via ilspycmd when needed, runs Roslyn, then embeds. Same long pipeline in one tool call — keep the MCP client open until it finishes. Needs OpenRouter and a built RoslynCodeGraph.exe.
This writes under PROJECT_DATABASES/my_game/.
Run the MCP server
python gamecode_rag_server.py
Cursor (mcp.json):
"gamecode-rag": {
"command": "C:/Python313/python.exe",
"args": [
"C:/path/to/gamecode-rag/gamecode_rag_server.py",
"--transport=stdio"
]
}
Grok (config.toml):
[mcp_servers.gamecode-rag]
command = 'C:\Python313\python.exe'
args = ['C:\path\to\gamecode-rag\gamecode_rag_server.py', "--transport=stdio"]
enabled = true
Pair with the suite
| Need | Server |
|---|---|
| Read Mono game code (offline index) | gamecode-rag (this repo) |
| Change the running game | bepinex-mcp |
| Read IL2CPP methods (static decompile) | il2cpp-decompiler |
Typical Mono flow: search here → find Player.TakeDamage → live patch with bepinex-mcp.
Layout
gamecode-rag/
gamecode_rag_server.py # MCP server
ingest_code_graph.py # CLI embed + call-graph save
embeddings_client.py # OpenRouter or local OpenAI-compatible embeds
tools/roslyn-parser/ # C# Roslyn → code_graph.json
PROJECT_DATABASES/ # your local indexes (gitignored)
requirements.txt
.env.example
Dockerfile
License
MIT — use at your own risk. You are responsible for how you obtain and index game code.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.