locopilot 0.1.3

An open-source, local-first, agentic coding assistant

🚀 Locopilot

Locopilot is an open-source, local-first, agentic coding assistant built for developers. It leverages local LLMs (via Ollama or vLLM), and advanced memory management using LangGraph, to automate, plan, and edit codebases—all inside an interactive shell.

• Private: All code and prompts stay on your machine.
• Agentic: Locopilot plans, edits, iterates, and manages your coding tasks.
• Interactive: Drop into a shell, enter tasks or slash commands, and steer the agent in real time.
• Memory-Efficient: Advanced memory compression via LangGraph for "infinite" context.
• Extensible: Change models, modes, and add custom tools or plugins on the fly.

Table of Contents

• Features
• How It Works
• Architecture
• Getting Started
• Usage: Interactive Shell & Commands
• Project Structure
• Extensibility & Roadmap
• Contributing
• License

✨ Features

• Local LLM Backend: Bring your own Ollama or vLLM server and code with any open-source LLM.
• LangGraph Agent Workflow: Plans, executes, edits, and compresses memory as a stateful, extensible graph.
• Interactive Shell/REPL: After init, drop into a chat-like agent terminal—just type coding tasks or slash commands.
• Slash Command Support: /model, /change-mode, /concise, /clear, /new, /end, /help, and more.
• Smart Memory Compression: Automatically summarizes previous context using the LLM itself, supporting ultra-long sessions.
• Configurable: Models, modes, and summarization thresholds are all runtime-editable.
• Pluggable Nodes: Add file tools, planning modules, git ops, and vector-based retrieval easily.
• (Planned) Git Integration: Auto-commit, rollback, and view code diffs per agent step.

⚡️ How It Works

1. Initialization

Run locopilot init in your project root.

• Locopilot checks Ollama/vLLM, prompts for backend/model, sets up .locopilot/config.yaml.
• You're dropped into an interactive agent shell (REPL).

2. Agentic Workflow (via LangGraph)

Each user input is parsed:

• Slash command (/model, etc.) → runs as a graph branch.
• Normal prompt (task) → plans, edits, summarizes via a workflow graph:

  User Task → [Planning Node] → [File Edit Node] → [Memory Summarizer Node] → (Repeat)
  

• Memory is managed with a LangGraph memory node—summarizing, chunking, and compressing context as needed.

3. Session Management

• Change models, modes, or reset memory on the fly with slash commands.
• All state (memory, model, mode) persists during the session.

🏗️ Architecture

Key components:

• CLI Layer: Typer-based CLI, launches shell (REPL), parses slash commands.
• LangGraph Workflow:
  • Nodes: Planning, file edit, summarization, slash command handler, etc.
  • Edges: Control session flow, branching between commands and prompts.
• LLM Backends:
  • Ollama: For running CodeLlama, DeepSeek, etc.
  • vLLM: OpenAI-compatible, GPU-powered.
• Memory Layer:
  • LangChain/LangGraph memory objects (buffer, summary, vector, hybrid).
  • Summarizes old context using the LLM to avoid hitting token/window limits.
• Config/Project Layer:
  • .locopilot/config.yaml stores model/backend/session preferences.

Stateful Graph Example:

               [User Input]
                      |
      +---------------+---------------+
      |                               |
 [Slash Command]              [Prompt/Task]
      |                               |
[Command Handler]   [Plan]->[Edit]->[Summarize]->[Memory]
      |                               |
     END                             Loop

🛠 Getting Started

Requirements

• Python 3.8+
• Ollama or vLLM running locally
• pip

Install Locopilot

Option 1: Install from PyPI (Recommended)

pip install locopilot

Option 2: Install from Source

git clone https://github.com/Ripan-Roy/locopilot-ai.git
cd locopilot-ai
pip install -e .

Start Your Local LLM

Ollama:

ollama serve
ollama pull codellama:latest

vLLM:

python -m vllm.entrypoints.openai.api_server --model <your-model>

Initialize and Enter the Agent Shell

locopilot init

This checks LLM backend, prompts for config, scans for project context, and launches the interactive shell.

🖥️ Usage: Interactive Shell & Commands

After init, Locopilot enters a shell where you can type prompts and commands:

Example Session

$ locopilot init
[✓] Ollama running. Model: codellama:latest
[✓] Project context initialized.

Locopilot Shell (mode: do):
> Add OAuth login to my Django app
[PLANNING] ...
[EDITING] ...
[MEMORY] ...

> /model
Current model: codellama:latest
Enter new model: deepseek-coder:latest
[✓] Model switched to deepseek-coder:latest

> /change-mode
Current mode: do
Available modes: do, refactor, explain, chat
Enter new mode: refactor
[✓] Mode set to refactor.

> Refactor the payment logic for clarity
...

> /concise
[✓] Context summarized and compressed.

> /clear
[✓] Session memory cleared.

> /new
[✓] New session started.

> /end
[✓] Session ended. Bye!

Supported Slash Commands

Command	Purpose
/model	Change LLM model/backend for current session
/change-mode	Switch between do, refactor, explain, chat modes
/clear	Clear all current context/memory
/new	Start a new session/project
/end	End the agent shell and exit
/concise	Force summarization/compression of current context
/help	Show help and command list

Anything not starting with / is treated as a task in the current mode!

🗂️ Project Structure

locopilot-ai/
├── __init__.py
├── cli.py                # CLI entrypoint, shell/repl logic
├── agent.py              # LangGraph workflow graph and nodes
├── agent_backup.py       # Backup of agent implementation
├── memory.py             # Session/context memory management
├── utils.py              # API, file, config helpers
├── connection.py         # Ollama/vLLM connection helpers
├── tests/
│   └── test_basic.py
├── dist/                 # Built distribution files
├── locopilot.egg-info/   # Package metadata
├── pyproject.toml
├── requirements.txt
├── setup.sh
├── README.md
├── LICENSE
└── env/                  # Virtual environment

🧠 Memory Management (with LangGraph)

• ConversationBufferMemory or ConversationSummaryBufferMemory is attached to the agent graph.
• As session context grows, old steps are summarized using the LLM and replaced in memory.
• This ensures Locopilot "remembers" key tasks, design decisions, and context for long sessions.
• Slash command /concise lets you summarize on demand.

⚡️ Extensibility & Roadmap

• Editor Plugins: VSCode, Vim, JetBrains, etc.
• Project-Aware RAG: Integrate vector DBs (Chroma, Qdrant) for smart codebase retrieval.
• (Planned) Git Integration: Auto-commit, diff, and rollback per step.
• Save/Load Sessions: /save, /load, /history commands.
• Custom Plugins/Nodes: Add your own LangGraph nodes for tools or workflows.
• Web/GUI Frontends: Same agent core, different interface.

🤝 Contributing

• Fork and PRs are welcome!
• Open issues for bugs or feature requests.
• For major features (graph nodes, memory backends), see CONTRIBUTING.md (coming soon).

📝 License

MIT License. Use, fork, and extend as you wish!

💡 Inspiration

Locopilot is inspired by Copilot, Claude Code, Dev-GPT, OpenDevin, and the emerging open-source agentic ecosystem—aiming to empower developers with private, supercharged, customizable AI tools.

🚦 Quickstart

# Install from PyPI
pip install locopilot

# Initialize in your project
locopilot init

# ... then just type your coding tasks and manage the session with slash commands!

Links:

• PyPI Package
• GitHub Repository