OpenLaoKe

Open-source terminal AI coding assistant with advanced automation, local model support, and intelligent supervision.

Overview

OpenLaoKe is a terminal-based AI coding assistant that supports 24+ AI providers and local GGUF models with zero API cost. Works with any model you choose — cloud API, local Ollama/LM Studio, or raw GGUF files.

Quick Start

pip install openlaoke
openlaoke

Key Features

• Interactive REPL — rich terminal UI, command history, smart autocomplete
• Multi-Provider — 24 cloud/local providers, any OpenAI-compatible endpoint
• Local GGUF Models — run any GGUF model locally via llama-cpp-python, zero API cost
• Ctrl+P Model Picker — instant provider/model switching overlay
• 30+ Tools — Read, Write, Edit, Glob, Grep, Bash, LSP, Git, WebSearch, compound tools
• MCP Support — connect to external Model Context Protocol servers
• Permission System — default / auto / bypass modes
• Session Persistence — auto-save and resume conversations
• Cost Tracking — real-time token usage and cost display
• 20+ Slash Commands — model switching, configuration, debugging
• Hook System — 15 extensible pre/post execution hooks
• Proxy Support — no proxy, system proxy, or custom proxy

Advanced

• HyperAuto — fully autonomous mode with self-improvement
• Task Supervision — automatic retry, completion verification, quality checking
• Model Assessment — 5-tier adaptive task decomposition
• Anti-AI Detection — human-appearing content with real citations
• Distilled Templates — 79 Q&A templates across 31 categories, multi-language triggers
• Skill System — 27+ YAML-based skills for specialized workflows
• Small Model Optimizations — type coercion, schema sanitization, read-loop prevention, output compression
• Fast Context Pruning — pure-algorithm compression (<5ms, no LLM call)
• Self-Reflection Tracker — auto-disables failing strategies, learns from outcomes
• Compound Tools — ReadAndPatch, FindAndRead, SearchAndRead reduce sequential calls
• Adaptive Router — auto-promotes to stronger models on failure
• Execution Traces — record/replay agent turns with regression test generation

Supported Providers

Free

Cloud

Local

Local GGUF Models

Run any GGUF model locally with llama-cpp-python. No API key, no network required.

pip install openlaoke
# llama-cpp-python is auto-detected; install manually if needed:
pip install llama-cpp-python

Download & Use Models

# Search ModelScope for any GGUF model
openlaoke model search llama

# Download a specific model
openlaoke model download unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF

# List downloaded models
openlaoke model list

# Remove a model
openlaoke model remove custom:unsloth-Llama-4-Scout-17B-16E-Instruct-GGUF

Configure

openlaoke --config
# Select "Built-in GGUF Model" → choose from downloaded models

Or edit ~/.openlaoke/config.json:

{
  "providers": {
    "active_provider": "local_builtin",
    "active_model": "custom:unsloth-Llama-4-Scout-17B-16E-Instruct-GGUF"
  }
}

Local Parameters

# In REPL
/localconfig n_ctx 32768
/localconfig temperature 0.5

Tools

File Operations

Read, Write, Edit, Glob, Grep, LS

Compound Tools

ReadAndPatch (read + edit in one call), FindAndRead (glob + read), SearchAndRead (grep + read)

Code Intelligence

LSP, Git, Bash (streaming), CodeRunner (sandboxed)

Web

WebSearch, WebFetch, WebBrowser (Playwright)

Task Management

TodoWrite, TaskKill, Batch, Agent (sub-agents)

Other

Notebook, Cron, Memory, REPL, Tmux, PowerShell

Slash Commands

CLI model management:

openlaoke model download [id]   # Download GGUF from ModelScope
openlaoke model list             # List downloaded models
openlaoke model search <query>   # Search ModelScope
openlaoke model remove <id>      # Remove model

Running Modes

openlaoke                              # Interactive TUI (default)
openlaoke --local                      # Local mode (atomic decomposition)
openlaoke web --host 0.0.0.0 --port 8080  # Web UI
openlaoke server                       # FastAPI backend (localhost:3000)
openlaoke "Write a Python script"      # Non-interactive
openlaoke --config                     # Configuration wizard
openlaoke --resume                     # Resume last session

Skill System

27+ YAML-based skills loaded from ~/.config/opencode/skills/:

Architecture

Configuration

~/.openlaoke/config.json:

{
  "providers": {
    "active_provider": "ollama",
    "active_model": "llama3.2",
    "providers": {
      "ollama": {
        "base_url": "http://localhost:11434/v1",
        "default_model": "llama3.2",
        "enabled": true
      },
      "openai": {
        "api_key_env": "OPENAI_API_KEY",
        "default_model": "gpt-4o",
        "enabled": false
      }
    }
  },
  "proxy_mode": "none",
  "max_tokens": 8192,
  "theme": "dark"
}

Environment Variables

Development

pip install -e ".[dev]"
ruff check . && ruff format .
mypy
pytest
pytest --cov

Acknowledgements

OpenLaoKe's architecture draws inspiration from several excellent open-source projects in the AI coding assistant space:

• nanobot — for the event-driven agent loop, multi-channel session model, AutoCompact, and Dream memory consolidation patterns
• smallcode — for the liquid tool-call parser, tool routing with category scoring, thinking budget control, plan-tracker with step anchoring, and read-before-write guard
• DeepSeek-Reasonix — for the transport-agnostic Controller pattern, cache-stable prefix with structured compaction, Previewer/PreviewChange tool interface, Config-driven provider/plugin registry, and plugin-based MCP client architecture
• OpenCode — for the full-screen TUI design, git-native workflow, session fork/branch checkpoint model, and the pragmatic approach to zero-config multi-model routing

License

GPLv3