A production-grade, local-first multi-agent AI coding assistant built on the Microsoft AutoGen framework. It leverages a local llama.cpp server (or equivalent OpenAI-compatible local endpoints) to run an autonomous crew of specialized agents that collaborate to plan, write, audit, test, and document code within your workspace.
The project is structured into three main layers:
├── backend/ # FastAPI Backend & Orchestrator
├── cli/ # Interactive Typer Command-line interface
├── frontend/ # Svelte/Vite/Tailwind Desktop UI
└── run_assistant.py # Main system runner entry point
Utilizes a Supervisor/Coordinator design pattern to orchestrate conversational turns in an AutoGen GroupChat across specialized roles:
- UserProxy: Represents you, manages local command execution, and executes tool calls.
- Coordinator (Supervisor): Audits progress, verifies outputs, and delegates milestones.
- Planner: Creates step-by-step roadmap documents detailing dependencies.
- Coder: Writes robust, modular, and self-documenting code.
- Reviewer / Tester / Debugger: Quality checks code, writes/runs tests, and acts on stderr tracebacks.
- Research / MCPAgent: Utilizes web search, semantic indexing, and Model Context Protocol (MCP) servers.
Indexes workspace files chunk-by-chunk using a local vector index. Provides semantic search capabilities using FAISS and embedding similarity calculations to allow agents to find code snippets.
Provides a rich dashboard experience:
- File Explorer & Code Editor: View workspace files recursively and edit them using an integrated Monaco Editor.
- Interactive Chat Logs: Streams agent thoughts (live thinking process) and messaging turns.
- Live Tool Running Trace: A console drawer logging tool actions (e.g. read/write files, shell commands).
- Imports Mapping Graph: Visualizes codebase import relationships.
Ensure you have a local llama.cpp server (or LM Studio) running on port 8082 serving a compatible model (e.g. gemma-4-12b-it-Q4_K_M.gguf).
- Base URL:
http://localhost:8082/v1 - Capabilities needed: The server must support chat completions with streaming enabled.
Install dependencies:
pip install -r requirements.txtInstall Node modules and compile the static bundle:
cd frontend
npm install
npm run buildThere are two primary ways to configure and customize your assistant:
Copy the provided .env.example to .env to configure your AI models:
- LLM_MODEL: The name of the model you are running (e.g.,
gemma-4-12b-it-Q4_K_M.gguf) - LLM_API_BASE: URL to your local server (e.g.,
http://localhost:8082/v1) - LLM_API_KEY: Your API key (use
not-neededfor local) - EMBEDDING_MODEL / URL / DIM: Configurations for the semantic codebase search.
You can easily customize all text, tooltips, and labels in the desktop UI without touching any Svelte code!
- Simply edit
frontend/src/labels.json. - The Vite server will automatically hot-reload your changes if you are running in development mode, or they will be bundled on your next
npm run build.
Use run_assistant.py to launch the assistant in different modes:
Starts the web backend on port 8000. If the frontend is built, it will serve the Web UI on http://localhost:8000.
python run_assistant.py serverInstructs the crew to execute a task immediately within the terminal:
python run_assistant.py task "Create a file index.html with simple hello world html"Starts an interactive terminal session where you can chat with the assistant or run slash commands:
python run_assistant.py cliInteractive CLI Commands:
/help- View available commands./plan <task>- Request the Planner to build a step-by-step roadmap./search <query>- Run a semantic code search across the workspace./git <args>- Run git status, diff, or log commands./review <file>- Audit a workspace file for bugs or improvements./exit- Quit the session.