v3.1.0 · Verifier · Skills · Doctor · LM Studio

Orchestration and
an interactive Pilot in one terminal

Latest GitHub release Latest PyPI version Release notes GPL-3.0-only OpenAI SDK → LM Studio CLI: lao
LAO · Pilot
 ██╗      █████╗  ██████╗ 
 ██║     ██╔══██╗██╔═══██╗
 ██║     ███████║██║   ██║
 ██║     ██╔══██║██║   ██║
 ███████╗██║  ██║╚██████╔╝
 ╚══════╝╚═╝  ╚═╝ ╚═════╝
     
 16:55:47 ▶ [Pilot] Entering Pilot Mode
 … weaving context from your workspace
    
  ❯ hi
    
 /help  /status  /resume  /project  ·  Enter send · Alt+Enter newline

Local AI Agent Orchestrator (LAO) runs a disciplined planner → coder → verifier → reviewer pipeline—and when work pauses, Pilot Mode gives you a chat similar to Claude Code or Cursor: tools, queue status, plans, and /resume back to autopilot. v3.1.0 adds mechanical verification, permissions, skills, hierarchical memory, optional MCP, and lao doctor. Multi-project? Use lao projects. State stays in SQLite; optional per-plan Git keeps history honest.

Quick start

Install and run in minutes

View on GitHub Install from PyPI
One-liner 
$ curl -fsSL https://lao.keyhan.info/install.sh | bash
pip 
$ pip install local-ai-agent-orchestrator
Run 
$ lao
  # choose "Initialize workspace" or "Pilot", or run `lao projects scan`
  $ lao health
  $ lao run
  $ lao pilot

LM Studio: Install the desktop app, start the local server, and load the models your factory.yaml keys refer to—LAO uses LM Studio’s HTTP load/unload API for memory-aware model switching (see README Prerequisites). The one-liner uses pipx if installed, otherwise pip install --user. Audit the script on GitHub before running. Existing project? Clone into ./MyRepo/ and use plans/MyRepo.md so LAO works directly in that codebase folder.

Capabilities

Built to squeeze serious work out of local models

✅
v3.1.0

Verifier & TODO ledger

A deterministic pass after coding checks files, syntax, and your published checklist—so the reviewer model is not wasted on obvious gaps.

🧩
v3.1.0

Skills & memory

Reusable Markdown skills, project LAO_MEMORY.md, user ~/.lao/MEMORY.md, and smarter context compaction for long sessions.

🔌
v3.1.0

Doctor · MCP · permissions

lao doctor for grouped health checks; optional MCP tool import; structured tool contracts with auto / plan_only / confirm modes.

🧭
v3.0.4+

Pilot Mode

Interactive agent with filesystem + shell tools, pipeline introspection, plan creation, /project switching, and guardrails when tools error repeatedly.

⚙️

Three-agent pipeline

Architect breaks plans into micro-tasks. Coder edits files with tools. Reviewer approves or sends structured rework—without bloated agent-framework scaffolding.

🗄️

SQLite backbone

WAL-mode queue survives crashes; retries, run logs, and token tallies stay queryable while jobs churn overnight.

🍎

Memory-aware swaps

Loads one big LLM at a time, unloads the rest, and waits for unified memory to settle—less swap thrash on Apple Silicon when models are tens of GB.

🔍

Semantic retrieval

Optional embedder (e.g. Nomic) narrows file context before coding so small local windows still hit the right files.

📦
New in v1.3

Git you can trust

Per-plan repos get LAO_PLAN.md, LAO_TASKS.json, and commits named lao(coder): … / lao(reviewer): … so progress is traceable.

🖥️
v2.0+

Unified interactive CLI

Home menu, guided init, model remap, orchestrator run, and Pilot share one coherent TTY language (Rich + prompt_toolkit) with clear next actions.

📂
v3.0.4+

Project registry

lao projects scan|use|list tracks known workspaces under ~/.lao/projects.json so you aren't stuck in a parent directory without factory.yaml.

Flow

From setup to Pilot and back

Open home

Run lao for status, grouped actions, and optional project scan.

Guided setup

lao init creates factory.yaml interactively and can bootstrap an initial plan.

Run pipeline

lao run executes planner/coder/reviewer with per-plan workspaces and unified TTY UI.

Pilot when idle

Chat, run tools, check status—or run lao pilot directly.

Resume autopilot

/resume or the resume tool returns control to the queue.

Pause and continue

SQLite + WAL: restart anytime; interrupted phases recover automatically.

In the terminal

See LAO v3.1.0

Grouped home menu for first-time operators; Pilot chat with scrollback, slash commands, and a live status strip—same build you get from pip install. New in v3.1.0: run lao doctor before long jobs.

LAO interactive home menu showing Initialize workspace, Pilot, and grouped other actions
Home menu — primary actions first, then Other actions, Exit last.
LAO Pilot Mode with LAO ASCII logo, chat transcript, and status bar
Pilot Mode — unified UI, tool rounds, token line, /resume to return to the pipeline.

Documentation

Learn more on GitHub

⚙️
Configuration

factory.yaml, paths, Git options

 🏗️
Architecture

Modules, queue, Git cadence

 📖
README

Full reference & changelog

 💬
Issues

Feedback & bug reports