Run a powerful AI coding agent on your own machine. Zero cloud. Zero cost. Zero data leaks. Powered by Omnigent (Databricks' open-source agent meta-harness) + Ollama.
OmniCoder Local is a minimal, production-ready AI coding agent that runs entirely on your hardware. It uses Omnigent — Databricks' newly open-sourced agent orchestration meta-harness — to wrap an Ollama-served language model into an agent capable of writing, refactoring, and debugging code — with no API keys, no subscriptions, and no internet connection required after setup.
Omnigent is a control plane that sits above existing AI agents. It provides a uniform API, governance controls (spend caps, tool-access policies, secure sandboxing), and seamless interoperability — so you can compose, govern, and share agents within a single unified system.
This project is the companion code for the YouTube tutorial: "Databricks Omnigent: The Secret Coding Agent Orchestration - Run Locally!"
⚠️ Clone the official Omnigent framework from Databricks first:# Clone the Omnigent meta-harness (required dependency) curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | shSee Step 3 in the Setup Guide below for full details.
| Layer | Technology |
|---|---|
| Agent Harness | Omnigent — Databricks open-source agent meta-framework (Apache 2.0) |
| Local LLM Runtime | Ollama — Run LLMs locally via a REST API |
| Model | qwen2.5-coder:1.5b — Fast, dedicated code generation model |
Omnigent's CLI depends on POSIX-only system modules (fcntl, SIGUSR1, tmux) that do not exist on Windows. Running omnigent in PowerShell or CMD will crash with an AttributeError or ModuleNotFoundError.
The fix: use WSL 2 (Windows Subsystem for Linux). It gives you a real Linux environment where Omnigent works perfectly, and your Windows files are accessible from inside it.
Windows native support is on Omnigent's roadmap. Track it at github.com/omnigent-ai/omnigent.
macOS / Linux users: skip to Step 2.
Open PowerShell as Administrator and run:
wsl --installThis installs WSL 2 with Ubuntu automatically. Restart your machine when prompted.
After restart, open the Ubuntu app from the Start menu. It will ask you to create a Linux username and password — do that, then keep working in that Ubuntu terminal for all steps below.
All remaining commands must be run inside the Ubuntu/WSL terminal, not PowerShell.
curl -fsSL https://ollama.com/install.sh | shThen pull the coding model:
ollama pull qwen2.5-coder:1.5bVerify Ollama is running:
curl http://localhost:11434/api/tagsYou should see a JSON list of your installed models.
The official installer handles Python, uv, and all dependencies automatically:
curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | shRestart your terminal (or run source ~/.bashrc), then verify:
omnigent --versionThe installer adds two aliases:
omnigentand the shorteromni. They are identical.
Run the interactive setup wizard:
omnigent setupWhen prompted:
- Choose "Add a credential"
- Select the openai-agents harness
- Choose Gateway as the credential type
- Set Base URL to:
http://localhost:11434/v1 - Set API key to any value (e.g.
ollama— Ollama does not validate keys) - Set the model to:
qwen2.5-coder:1.5b
omnigent run local_coder_agent.yaml --prompt "Build me a responsive landing page for a SaaS product."Save output to a file:
omnigent run local_coder_agent.yaml --prompt "Build a dark-mode todo app in a single HTML file." --output todo_app.html# 1. Confirm Ollama is live
curl http://localhost:11434/api/tags
# 2. Confirm Omnigent CLI is installed
omnigent --version
# 3. Run the agent with a quick prompt
omnigent run local_coder_agent.yaml --prompt "Write a CSS button with a hover animation."# local_coder_agent.yaml
name: local_hacker # Unique agent identifier used by Omnigent
prompt: >
You are an elite frontend developer.
Your only job is to write single-file,
responsive, beautiful HTML/CSS/JS applications.
Output raw code without markdown wrappers if possible.
executor:
harness: openai-agents # Omnigent's OpenAI-compatible harness — works with OllamaKey insight: Ollama exposes an OpenAI-compatible REST API at /v1. Omnigent's openai-agents harness speaks this protocol natively. The gateway URL and model are configured once via omnigent setup and reused across all agents — no credentials need to be hardcoded in the YAML.
- Rapid Prototyping — Generate full UI components and landing pages from a single sentence.
- Offline Development — Code with AI assistance in air-gapped or low-connectivity environments.
- Privacy-First Workflows — Keep proprietary code and internal tooling off cloud servers.
- CI/CD Integration — Embed the agent into local build pipelines for automated code generation steps.
- Learning & Exploration — Experiment with prompt engineering and agent design without any API costs.
- File-aware context: Feed the agent your entire project directory so it can refactor across multiple files.
- Tool use / function calling: Enable the agent to read/write files, run shell commands, and browse documentation.
- Multi-agent pipelines: Chain a
planneragent and acoderagent using Omnigent's orchestration layer. - Web UI: Add a lightweight Gradio or Streamlit frontend for non-terminal users.
- Model benchmarking: Auto-run a prompt suite against multiple local models and score output quality.
RayCodes_Omnigent/
├── README.md # This file
├── local_coder_agent.yaml # Omnigent agent definition
└── .gitignore # Excludes large model files
MIT — Free to use, fork, and ship.
Built with ❤️ for the open-source AI community.
Keywords: Databricks Omnigent, AI agent orchestration, local AI coding agent, Ollama local LLM, run AI agent locally, free AI coding agent, qwen2.5-coder, agent meta-harness, open source AI agent, Omnigent tutorial, AI coding offline, LLM orchestration, agentic AI workflow, coding agent WSL, multi-agent orchestration, Databricks open source, AI agent framework 2025, local code generation, privacy AI coding, zero cloud AI