📺 Watch the full tutorial on YouTube

Xiaomi’s MiMo Code Just Dropped: The Ultimate Coding AI Agent!

[ 🟢 100% Air-Gapped Offline ]  •  [ 🤖 Local Model: qwen2.5-coder:1.5b ]  •  [ 💻 OS: Windows PowerShell ]

An enterprise-ready, fully air-gapped offline AI developer stack designed to orchestrate code modifications, track agent execution logs, and maintain persistent session checkpoints locally on Windows environments. Powered by Ollama's efficient local models and MiMo-Code's memory-aware agent framework.

---

🛠️ System Architecture

┌─────────────────────────────────────────────────────────┐
│                    Local Workstation                    │
│                                                         │
│   ┌──────────────┐      HTTP API      ┌─────────────┐   │
│   │  MiMo-Code   │ ────────────────>  │   Ollama    │   │
│   │  CLI / TUI   │ <────────────────  │ Local Daemon│   │
│   └──────┬───────┘    (Port 11434)    └──────┬──────┘   │
│          │                                   │          │
│          │ File System                       │ Model    │
│          ▼ Changes                           ▼ Weights  │
│   ┌──────────────┐                    ┌─────────────┐   │
│   │  Workspace   │                    │ qwen2.5-    │   │
│   │ (app.py, etc)│                    │ coder:1.5b  │   │
│   └──────────────┘                    └─────────────┘   │
└─────────────────────────────────────────────────────────┘

---

📋 Project Overview

This workspace serves as a dedicated local environment demonstrating the deployment of an offline developer assistant. By utilizing MiMo-Code and a local Ollama inference server running qwen2.5-coder:1.5b, developers can execute multi-turn coding plans, auto-reconstruct context, and debug scripts with zero telemetry or external network calls.

Key architectural components include:

1. 🤖 Local LLM Engine: Self-hosted Ollama server providing fast code completion and editing instructions.
2. 🔌 Context-Aware CLI Agent: MiMo-Code CLI running on a local port with configuration mappings.
3. 📊 Visual Live Dashboard: An interactive Streamlit dashboard allowing users to submit TUI prompts directly to the local model, watch agent execution logs, and monitor file modifications side-by-side.

---

💻 Core Tech Stack

Component	Technology	Description
Model Host	Ollama	Windows native offline LLM server service
Local LLM	qwen2.5-coder:1.5b	<1.2GB footprint optimized coding assistant
Orchestrator	MiMo-Code	Memory-aware CLI/TUI agent (fork of OpenCode)
Runtime	Python 3.x	Code execution & debugging environment
Dashboard	Streamlit (Python)	Interactive live Web UI with real-time model connections

---

🚀 PowerShell Installation Guide for Windows

Execute the following PowerShell commands in your terminal to bootstrap the environment.

1️⃣ Install Ollama and Pre-cache Model

# Download and install Ollama for Windows
Start-Process -FilePath "https://ollama.com/download/OllamaSetup.exe" -ArgumentList "/silent" -Wait

# Verify service status and pull the lightweight coding model
ollama run qwen2.5-coder:1.5b "Hello"

2️⃣ Install MiMo-Code CLI Agent

# Install MiMo-Code CLI globally
npm install -g @mimo-ai/cli

Important

Ensure the Ollama service is running locally (http://localhost:11434) before initiating the dashboard or the MiMo-Code CLI interface.

---

🏃‍♂️ Running the Live Demo

You can execute the code repair demo via two different interfaces: the Streamlit Web Dashboard or the Terminal CLI.

💻 Path A: Streamlit Live Interactive Dashboard & TUI

Launch the dashboard to submit prompts and watch the agent perform code modifications in real time:

1. Launch the Dashboard:

   streamlit run dashboard.py

2. Submit a TUI Prompt:
   • Open your browser at http://localhost:8501.
   • Enter your prompt (e.g., Fix the list mutation bug in app.py) in the input field.
   • The dashboard calls Ollama, updates app.py, updates the live log feed, and displays the code changes side-by-side.
3. Verify the Repaired Code:

   python app.py

🐚 Path B: Direct Terminal Commands (MiMo-Code CLI)

Use the MiMo-Code CLI TUI directly in the workspace folder:

1. Start the MiMo Agent TUI:

   mimo

2. Submit the Prompt inside the Agent Interface:

   audit app.py, detect the list-mutation bug, and write the corrected clean_pending_records function directly back into app.py
   

3. Verify Output:

   python app.py

Expected output after correction: Result: [{'id': 3, 'status': 'active'}] (The pending items are correctly removed without skipping any adjacent elements.)

---

📂 Workspace Directory Breakdown

• 📄 README.md This configuration manual and deployment architecture overview.
• ⚙️ config.json Explicit mapping file containing API endpoint routing, model details, and orchestration flags targeting the local Ollama daemon.
• 🐍 app.py A local Python data processing script containing deliberate data-loop bottlenecks and bugs designed to serve as evaluation targets.
• 📊 dashboard.py An interactive, real-time Streamlit dashboard script that connects directly to Ollama to test TUI prompting and visualize code changes.

---

🏢 Enterprise Deployment Use Cases

1. 🔒 Air-Gapped Government Networks Secure coding environments for defense contractors and agencies where external internet access is prohibited due to high confidentiality standards.
2. 🏦 Proprietary Banking Codebases Financial institutions requiring zero-leakage policies on source code, ensuring proprietary transaction algorithms never traverse public APIs.
3. 🚢 High-Speed Developer Workstations Workstations on low-bandwidth networks (e.g., ships, remote drilling platforms) needing immediate, zero-latency autocomplete capabilities.
4. 🏥 On-Premises Compliance Sandboxes Healthcare organizations handling sensitive patient data APIs requiring HIPAA compliance and localized software patch generation.
5. 🔬 Standardized AI Training labs Offline academic or corporate labs running automated model evaluations against local base models in isolated network segments.

---

🎯 Future Evolutionary Milestones

• Local Vector-Based RAG Ingestion: Integrating local vector stores (e.g., Turbovec/SQLite FTS5) directly in the workspace to allow repository queries.
• Multi-Model Orchestration: Dynamic switching between coding models and reasoning models via a local router.
• Structured Live Diffs: Visual side-by-side terminal-native diff rendering on file system change triggers.
• Offline Skill Synthesis: Automatic packaging of repetitive developer steps into custom skill subagents.
• Windows-Native Hotkey Integration: Direct keyboard hook shortcuts to trigger terminal-agent actions from standard editors.

---

🔑 Keywords & SEO tags

MiMo Code Xiaomi MiMoCode Offline AI Agent Ollama qwen2.5-coder Local LLM Developer Air-Gapped AI Coding Windows PowerShell Developer Stack OpenCode Fork AI Coding Assistant Private AI agent SQLite FTS5 FTS Autopilot Coding