A model-agnostic agentic coding CLI tool that works with local LLMs via Ollama. OpenCoder enables AI-assisted coding directly in your terminal with file operations, bash commands, and intelligent code analysis.

• Agentic Workflow: AI that can read, write, and edit files autonomously
• Tool Calling: Built-in tools for file operations, bash commands, and glob patterns
• Model Agnostic: Works with any Ollama-compatible model
• Planning System: Break down complex tasks into executable steps
• Interactive REPL: Conversational interface with slash commands

• Node.js >= 18.0.0
• Ollama running locally (or accessible server)

# Clone the repository
git clone https://github.com/YOUR_USERNAME/opencoder.git
cd opencoder

# Install dependencies
npm install

# Build the project
npm run build

# Link globally (optional)
npm link

First, install Ollama from ollama.ai, then pull a model:

# Pull a recommended model
ollama pull deepseek-r1:8b

# Start Ollama (if not already running)
ollama serve

# If installed globally
opencoder

# Or run directly
npm start

# Or with specific options
opencoder --model deepseek-r1:8b --url http://localhost:11434

Once in the REPL, you can ask the AI to help with coding tasks:

You: Read the main.py file and explain what it does
You: Create a new function that validates email addresses
You: Find all TODO comments in the src directory

OpenCoder uses prompt-based tool calling since most local models don't have native function calling. Results vary by model:

Tip: Start with deepseek-r1:8b for the best balance of speed and capability. Experiment with larger models for complex reasoning tasks.

Configuration is loaded in priority order (highest to lowest):

opencoder --model deepseek-r1:32b --url http://192.168.1.100:11434

export OPENCODER_MODEL=deepseek-r1:8b
export OPENCODER_BASE_URL=http://localhost:11434
export OPENCODER_PROVIDER=ollama

Create ~/.opencoder/config.json:

{
  "provider": "ollama",
  "model": "deepseek-r1:8b",
  "baseUrl": "http://localhost:11434",
  "timeout": 300000
}

• Provider: ollama
• Model: deepseek-r1:8b
• URL: http://localhost:11434
• Timeout: 300000ms (5 minutes)

opencoder [options]

Options:
  -m, --model <model>     Model to use (default: deepseek-r1:8b)
  -p, --provider <name>   AI provider (default: ollama)
  -u, --url <url>         Base URL for API (default: http://localhost:11434)
  -h, --help              Display help

Inside the REPL:

The AI has access to these tools:

src/
├── index.ts          # CLI entry point
├── agent/            # Core agentic loop
│   ├── agent.ts      # Main agent with tool execution
│   ├── context.ts    # Conversation management
│   └── parser.ts     # Tool call extraction
├── providers/        # AI provider abstraction
│   ├── base.ts       # Provider interface
│   └── ollama.ts     # Ollama implementation
├── tools/            # Agent capabilities
│   ├── read.ts       # Read files
│   ├── write.ts      # Write files
│   ├── edit.ts       # Edit files
│   ├── bash.ts       # Execute commands
│   └── glob.ts       # File patterns
├── cli/              # REPL interface
└── planning/         # Task planning system

# Watch mode for development
npm run dev

# Build
npm run build

# Run
npm start

Extend the AIProvider base class in src/providers/:

import { AIProvider, Message, ChatResponse } from './base.js';

export class MyProvider extends AIProvider {
  async chat(messages: Message[]): Promise<ChatResponse> {
    // Implement your provider logic
  }

  async checkConnection(): Promise<boolean> {
    // Verify connection to the service
  }
}

Make sure Ollama is running:

ollama serve

Pull the model first:

ollama pull deepseek-r1:8b

• Try a smaller model (8b vs 32b)
• Increase timeout in config
• Check system resources

Some models don't follow tool-calling instructions well. Try:

1. Using deepseek-r1:8b (best tool compliance)
2. Being more explicit in your prompts
3. Breaking complex tasks into smaller steps

MIT

Contributions are welcome! Please feel free to submit issues and pull requests.