Delia is an MCP server that enhances AI coding assistants with persistent learning, semantic code intelligence, and GPU-powered task execution.
Core Capabilities:
GPU-Powered Task Execution: Delegates code review, generation, analysis, planning, and summarization tasks to local/remote GPUs with intelligent 3-tier model routing (quick/coder/moe) based on content size and task type
Deep Reasoning: Performs complex multi-step problem-solving, architectural design, and debugging strategies using extended thinking mode with configurable depth levels
Parallel Processing: Executes multiple tasks concurrently across all available GPUs for maximum throughput
Semantic Code Intelligence: Provides LSP-like navigation including finding references, going to definitions, and renaming symbols through codebase indexing
Persistent Learning: Stores project patterns (playbooks), knowledge (memories), and framework guidance (profiles) in ChromaDB; extracts insights from completed tasks to continuously improve playbooks
Backend Management: Monitors GPU backend health, tracks usage stats and cost savings, manages model queues and VRAM, and allows dynamic switching of backends and models at runtime without restarting
Intelligent Routing: Automatically selects backends based on content size (>32K tokens routes to largest context window), prefers local GPUs for lower latency with fallback to remote, and includes load balancing with circuit breaker protection
Project Integration: Initializes projects with tech stack-tailored playbooks and integrates with AI coding assistants via MCP protocol
Integrates with GitHub Copilot CLI to provide access to local LLM capabilities through MCP tools for task delegation and reasoning.
Supports Google Gemini as an optional cloud fallback backend for task delegation and reasoning when local models are unavailable.
Provides integration with Ollama for running local LLMs, including automatic model selection, context-aware routing, and support for multiple model tiers (quick, coder, MoE, thinking).
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@DeliaInitialize this project with Delia"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Delia
MCP server that adds persistent learning and semantic code intelligence to AI coding assistants.
What It Does
Playbooks - Per-project patterns learned over time, indexed in ChromaDB for semantic retrieval
Memories - Persistent knowledge (markdown), searchable via embeddings
Profiles - Framework-specific guidance, semantically matched to your task
Code Index - Codebase summaries and symbols indexed for intelligent navigation
LSP Tools - Semantic code navigation: find references, go to definition, rename symbols
Learning Loop - Extracts insights from completed tasks and updates playbooks
All knowledge is stored in .delia/chroma/ for fast semantic search.
Full Documentation | Quick Start | Tool Reference
Quick Start
# 1. Clone and install
git clone https://github.com/zbrdc/delia.git
cd delia
uv sync
# 2. Start HTTP server (recommended for multi-project)
uv run delia run -t http --port 8765
# 3. Initialize your project (from project directory)
cd ~/your-project
uv run --directory ~/git/delia delia init-projectComplete Setup Guide
Step 1: Install Delia
git clone https://github.com/zbrdc/delia.git
cd delia
uv syncStep 2: Configure MCP Clients
Auto-detect and configure all supported AI clients:
uv run delia installOr install to a specific client:
uv run delia install claude # Claude Code
uv run delia install cursor # Cursor
uv run delia install vscode # VS CodeList available clients:
uv run delia install --listStep 3: Start the Server
Option A: HTTP Transport (Recommended)
Best for multi-project setups. One server handles all projects.
uv run delia run -t http --port 8765Add to each project's .mcp.json:
{
"mcpServers": {
"delia": {
"type": "http",
"url": "http://localhost:8765/mcp"
}
}
}Note: HTTP servers won't appear in Claude Code's /mcp list, but tools work normally.
Option B: stdio Transport
Per-project server, managed by the AI client. Shows in /mcp list.
{
"mcpServers": {
"delia": {
"command": "uv",
"args": ["--directory", "/path/to/delia", "run", "delia", "serve"]
}
}
}Step 4: Initialize Your Project
Option A: Via MCP (Recommended)
Let the AI agent initialize the project - it handles summarization:
# In Claude Code or Cursor, just ask:
"Initialize this project with Delia"
# Or use the MCP tool directly:
project(action="init", path="/path/to/your-project")Option B: Via CLI (requires Ollama)
If you have Ollama running locally with a model:
cd ~/your-project
uv run --directory /path/to/delia delia init-projectThis creates .delia/ with playbooks tailored to your tech stack.
Step 5: Verify Setup
uv run delia doctorUsage
The AI assistant calls these tools:
auto_context("implement user auth") # Load relevant patterns
[work on the task]
complete_task(success=True, bullets_applied=["id1"]) # Record feedbackProject Structure
your-project/
├── .delia/
│ ├── chroma/ # Vector database (primary storage)
│ ├── playbooks/ # Learned patterns (JSON, indexed to ChromaDB)
│ ├── memories/ # Persistent knowledge (Markdown, indexed to ChromaDB)
│ └── profiles/ # Framework guides (Markdown, indexed to ChromaDB)
└── CLAUDE.md # Instructions for AI assistantsCLI Commands
delia run -t http # Start HTTP server (MCP)
delia serve # Start stdio server (MCP)
delia doctor # Health check
delia init-project # Initialize project (requires Ollama)
delia chat # Interactive chat (requires Ollama)
delia agent "task" # Single-shot task (requires Ollama)Configuration
Embeddings (for semantic search)
Create ~/.delia/.env:
DELIA_VOYAGE_API_KEY=your-key-hereFallback options (no API key needed):
Ollama - Run
ollama pull mxbai-embed-largeSentence Transformers - CPU fallback, works offline
LLM Backends (for CLI features)
For init-project, chat, agent commands, configure backends in ~/.delia/settings.json:
{
"backends": [{
"name": "ollama-local",
"url": "http://localhost:11434",
"model": "llama3.2"
}]
}Requirements
Python 3.11+
uv (package manager)
Ollama (optional, for CLI LLM features - not needed if using MCP only)
License
GPL-3.0
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.