Coding Agent

A local AI coding assistant powered by Ollama, built with TypeScript. Supports agentic workflows — reads files, writes code, runs commands, and iterates until the task is done.

Features

• Agentic loop — model autonomously calls tools until task is complete
• 6 built-in tools — read_file, write_file, str_replace, delete_file, list_dir, run_command, search_code
• Safety guards — path traversal protection, readonly file guard, dangerous command blocking, file size limit
• Diff view — shows changes before applying, requires confirmation for destructive operations
• Streaming output — token-by-token display as model thinks
• Slash commands — /reset, /history, /help, /exit
• Fully local — no API keys, no internet required after model download

Stack

• Runtime: Node.js 20+ with TypeScript
• LLM: Ollama (llama3.1:8b or qwen2.5-coder:7b)
• CLI: Chalk, Ora, Inquirer, Commander

Prerequisites

• Node.js 20+
• Ollama

Setup

# Clone the repo
git clone https://github.com/phuoctrung-ppt/coding-agent
cd coding-agent

# Install dependencies
npm install

# Pull the model
ollama pull llama3.1:8b
ollama pull nomic-embed-text

# Create your model with system prompt
ollama create fullstack-dev -f Modelfile

# Copy and configure environment
cp .env.example .env

Edit .env:

OLLAMA_MODEL=fullstack-dev
PROJECT_DIR=./path/to/your/project   # point to the codebase you want to work with

Usage

npm run dev

You: list all files in src and explain the project structure
You: add a soft delete method to UserService
You: find all places where any type is used and fix them

How it works

The agent follows a simple loop:

1. User sends a message
2. Model decides: respond with text, or call a tool
3. If tool call → agent executes it and feeds result back to model
4. Loop repeats until model returns plain text (task done)

This is the same pattern used by Claude Code, Cursor, and most AI coding assistants — the difference is the underlying model.

Upgrading to a stronger model

The agent uses the OpenAI-compatible API format. To use a more capable model, update OLLAMA_MODEL in .env or swap the Ollama client for the Anthropic/OpenAI SDK — no other code changes needed.

Project structure

src/
├── agent/
│   ├── core.ts       # agentic loop + tool routing
│   ├── tools.ts      # file and shell tool implementations
│   └── guards.ts     # safety checks for all tool calls
├── cli/
│   └── app.ts        # terminal interface, streaming, diff view
├── config/
│   └── settings.ts   # environment config
└── index.ts          # entry point

License

MIT