ProdE MCP Server for Contextual Codebase AI Assistance

Overview

ProdE MCP Server implements the Model Context Protocol (MCP) to provide contextual codebase understanding and AI assistance across multiple projects. It centralizes code indexing, context retrieval, and model-aware snippets so language models can access relevant code, docs, and config when answering developer queries, generating code, or reviewing changes. The server is designed to sit between your code repositories and your LLMs/AI tools, surfacing precise context and reducing hallucinations.

For teams working across many repositories, ProdE simplifies cross-repo search, context-aware completions, and automated code-review workflows by maintaining searchable context stores, handling incremental updates, and exposing a simple HTTP API for queries. It supports common embedding providers and vector stores so you can plug it into your existing AI toolchain.

Features

• Multi-project indexing and context stores
• Incremental file watching and delta updates
• Retrieval of code, docs, and config with model-aware chunking
• Pluggable embedding providers and vector store connectors
• HTTP API supporting query, project management, and ingestion
• Configurable retrieval strategies (top-k, score thresholds, time window)
• Simple authentication and health/metrics endpoints
• Designed to reduce LLM hallucination by returning precise source fragments

Installation / Configuration

Clone the repository and run with Docker or locally. Example commands below assume a Unix-like shell.

Clone:

git clone https://github.com/CuriousBox-AI/ProdE-mcp.git
cd ProdE-mcp

Run with Docker Compose (recommended for production-like setup):

docker-compose up --build

Run locally (node/npm example):

# install dependencies
npm install

# build and start
npm run build
npm start

Environment variables (example .env):

# Server
MCP_PORT=8080
MCP_HOST=0.0.0.0

# Embeddings provider (example OpenAI)
EMBEDDINGS_PROVIDER=openai
OPENAI_API_KEY=sk-...

# Vector store (example)
VECTOR_STORE=faiss
VECTOR_STORE_PATH=./data/faiss

# Security
MCP_API_KEY=changeme

Project configuration (example projects.yaml):

projects:
  - id: payments
    path: /home/ci/repos/payments
    exclude:
      - node_modules
      - .git
    languages: ["ts","js","sql"]
  - id: frontend
    repo: https://github.com/org/frontend.git
    branch: main

After configuration, use the API to register projects or trigger an initial index.

Available Resources

• HTTP API endpoints for programmatic integration (see table below)
• Connectors: embedding providers (OpenAI, Hugging Face), vector stores (FAISS, Milvus, Pinecone) — configure via env
• Ingestion adapters: local filesystem, Git repositories, archive uploads
• SDK examples and integrations (client libraries, CLI) in the repo
• Logs, metrics, and health endpoints to monitor indexing and retrieval performance

API overview

Example query request:

curl -X POST "http://localhost:8080/v1/query" \
  -H "Authorization: Bearer $MCP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "project": "payments",
    "prompt": "How does the refunds workflow handle idempotency?",
    "top_k": 5
  }'

Typical response (JSON):

{
  "query_id": "q_123",
  "results": [
    {
      "score": 0.93,
      "source": "payments/src/refunds/service.ts:42-89",
      "text": "function processRefund(...) { ... }"
    },
    ...
  ]
}

Use Cases

• Context-aware code completions: Provide an LLM with the exact functions, types, and configuration files relevant to the editing buffer so completions and suggestions reference real code.
• Pull request summarization and review: Retrieve changed files and related code snippets to generate accurate PR summaries and automated review comments.
• Cross-repo impact analysis: Query multiple projects to find where a function or configuration is referenced, helping with safe refactors.
• On-call debugging: Quickly surface stack-trace-relevant source snippets and recent deployments/config so incident responders can triage faster.
• Developer onboarding: New team members can query behavior and architecture across several repositories without needing full local builds or manual search.

Getting Help and Next Steps

• Browse the source, issues, and examples on the GitHub repo: https://github.com/CuriousBox-AI/ProdE-mcp
• Start by registering one project, running an initial index, and making a few /v1/query calls to validate retrieval quality.
• If integrating with your LLM pipeline, experiment with different embedding providers and top_k values to balance context size and precision.

This server is intended to be integrated into developer tools, CI/CD workflows, and LLM prompts to ground AI-assisted development with reliable, project-specific context.