Woreda

The Ollama-first AI coding agent built for the terminal.

Forked from OpenCode to focus exclusively on local Ollama models with tool support.

Philosophy

Woreda is 100% local-first. No cloud APIs, no billing, no authentication. Just you, your powerful machine, and open-source models running locally via Ollama.

Why local-only?

• Privacy: Your code never leaves your machine
• Cost: $0.00 per token, forever
• Speed: No network latency, just raw local inference
• Control: Full context window management, no provider limits
• Focus: Built for code, not chat

We target Ollama models with tool support - specifically optimized for coding tasks like qwen2.5-coder, deepseek-coder-v2, and llama3.2.

Installation

Prerequisites

1. Install Ollama: https://ollama.com/download
2. Pull a tool-capable model:

   ollama pull qwen2.5-coder        # Recommended: Best for code
   ollama pull deepseek-coder-v2    # Alternative: Great for reasoning
   ollama pull llama3.2             # Alternative: Fast and capable
   

Install Woreda

# From source (for now)
git clone https://github.com/samifouad/woreda.git
cd woreda
bun install
bun run build
bun link

# Use it
woreda spawn

Supported Models

Woreda only shows Ollama models with tool support. These models can execute bash commands, edit files, and use the full agent toolkit.

Recommended:

• qwen2.5-coder - Best overall for code generation
• deepseek-coder-v2 - Excellent reasoning capabilities
• llama3.2 - Fast and lightweight (3B/1B variants available)

Also supported:

• llama3.1 - Strong general capabilities
• mistral-nemo - Good balance of speed and capability
• command-r - Cohere's code model
• firefunction-v2 - Fireworks' function-calling specialist

Model Requirements

To appear in Woreda, a model must:

1. Be available in your local Ollama instance
2. Support tool/function calling
3. Match one of the known tool-capable model families

Features

Agent System

Switch between agents with the Tab key:

• build (default) - Full access agent for development work
• plan - Read-only agent for analysis and exploration
  • Denies file edits by default
  • Asks permission before running bash commands
  • Perfect for exploring unfamiliar codebases

Tools

All the standard tools you need:

• bash - Execute shell commands (with tree-sitter safety parsing)
• edit - Intelligent file editing with diffs
• read - Smart file reading
• write - File creation
• grep - Pattern-based code search
• glob - File pattern matching
• task - Multi-step task execution
• mcp - Model Context Protocol support

MCP Support

Woreda keeps full MCP (Model Context Protocol) support. Connect external tools and servers:

// .opencode/opencode.jsonc
{
  "mcp": {
    "servers": {
      "filesystem": {
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/files"]
      }
    }
  }
}

Context Window Management

One of Woreda's core focuses is intelligent context window management for local models:

• Auto-detects model context limits (32K, 64K, 128K)
• Smart message compaction
• Session snapshots and revert
• Optimized for tool-heavy workflows

Configuration

Woreda uses the same config format as OpenCode for compatibility:

// .opencode/opencode.jsonc (or ~/.opencode/opencode.jsonc)
{
  "model": "ollama/qwen2.5-coder:latest",
  "provider": {
    "ollama": {
      "api": "http://localhost:11434/v1",  // Override if needed
      "models": {
        "custom-model:latest": {
          "name": "My Custom Model",
          "tool_call": true  // Mark as tool-capable
        }
      }
    }
  }
}

Environment Variables

• OLLAMA_HOST - Override Ollama API location (default: http://localhost:11434)

Usage

# Start TUI
woreda spawn

# Run a one-off command
woreda run "refactor this function to use async/await"

# Attach to running server
woreda tui attach

Development

# Clone and setup
git clone https://github.com/samifouad/woreda.git
cd woreda
bun install

# Run in dev mode
bun dev

# Type checking
bun typecheck

# Tests
bun test

# Build
bun run build

Roadmap

Core Focus: Local Model Excellence

• Enhanced context window management
• Model-specific prompt optimization
• Advanced agent collaboration
• Session branching and merging
• Intelligent tool selection per model

Ollama Integration

• Auto-pull missing models
• Model performance benchmarking
• Quantization recommendations
• num_ctx auto-tuning
• Model switching based on task

Developer Experience

• Rich TUI improvements (inspired by neovim)
• Better error handling for local models
• Streaming performance optimization
• Offline-first architecture

FAQ

Why fork OpenCode?

OpenCode is excellent but deliberately multi-provider. It treats Ollama as a second-class citizen requiring manual JSON configuration. Woreda inverts this: Ollama is the only citizen, and we optimize everything around local model performance.

Why Ollama only?

Focus breeds excellence. By targeting only Ollama, we can:

• Optimize context window management for local constraints
• Build features specific to local model capabilities
• Remove all authentication, billing, and cloud complexity
• Provide a superior local-first developer experience

What about cloud models?

They're great! But they're also well-served by existing tools (including OpenCode, Claude Code, Cursor, etc.). Woreda carves out a niche: local-only, privacy-first, cost-free coding.

Can I still use MCP?

Absolutely! MCP is provider-agnostic and works great with local models.

Will you add cloud providers back?

No. That's the whole point. Use OpenCode if you want multi-provider support.

Credits

Forked from OpenCode by the SST team. Massive respect for their work building the foundation.

Changes:

• Ripped out all cloud providers (Anthropic, OpenAI, Google, AWS, Azure)
• Ripped out authentication system
• Focused exclusively on Ollama models with tool support
• TUI-only (removed web console and desktop app)
• Renamed to Woreda (Ethiopian for "district" - representing local community)

License

MIT (inherited from OpenCode)