AI Company

Turn Claude Code into your personal embedded + full-stack dev team.

Any idea → working software/firmware.

👑 Team Structure

Chairman (You)  →  CEO (Claude AI)  →  18 Specialized Agents
    🧑‍💼                  🤖                   👥
 Set direction     Autonomous execution    Write firmware,
 Approve strategy  Task scheduling         design hardware,
 Final decisions   Pipeline management     review code, test,
                   Research scanning       deploy, document

Role	Who	Responsibility
🧑‍💼 Chairman	You	Set the vision. Approve the big decisions. The system works for you.
🤖 CEO	Claude AI	Runs the company 24/7. Picks up tasks, manages pipelines, scans for new tech. Zero prompts needed.
👥 18 Agents	Specialized AI roles	Firmware, hardware, Linux, IoT, frontend, backend, DevOps, security, QA, research...

💡 You're not the operator. You're the owner. The AI is your CEO.

📖 Read this in: 繁體中文

🎯 Core Capabilities

Capability	Description
🧠 Idea → Software	Complete pipeline from vague idea to deployment
🔌 Hardware Bridge	Auto-detect dev boards, flash firmware, monitor serial
👥 18 AI Agents	Embedded × 6 + Software × 5 + Management × 5 + Special × 2
📋 5 Pipelines	Firmware / Linux / Fullstack / Prototype / Research
💾 Knowledge Base	M5Stack, ESP32 pin definitions and code patterns built-in
🔒 Zero External Cost	No external API calls, fully local operation
⏱ Time Tracking	Automatic per-task elapsed time, pause/resume on status changes, per-agent breakdown
🪙 Token Tracking	Per-agent token consumption tracking, cumulative totals across projects
🧬 Self-Evolution	Auto-detect slow tasks, create failure records, trigger improvement cycles

🚀 Quick Start

Installation

# 1. Install uv (if you haven't already)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Clone
git clone https://github.com/ian0318git/AI-company.git
cd AI-company

# 3. Sync dependencies
uv sync

# 4. Install into Claude Code
uv run python install.py

# 5. Restart Claude Code

Your First Project

Just say this inside Claude Code:

"I want to build a temperature and humidity monitor for M5Stack Core S3 that displays readings on the LCD and uploads data to the cloud via WiFi"

The system will automatically:

1. Idea Refiner — Refine the requirements
2. Pipeline — Create an embedded-firmware workflow
3. Firmware Engineer — Write ESP32-S3 code
4. Hardware Engineer — Plan pin assignments
5. Compile → Flash to your M5Stack

After the pipeline is created, don't wait for tasks to run by themselves. Open the Dashboard → Idea Inbox → click "Start Auto Schedule" (top bar) → watch tasks execute in real time.

💡 Common mistake: Users submit an idea, start a pipeline, and wonder why nothing happens. Tasks start in todo state — you must kick off the scheduler or click individual tasks to begin execution.

🖥️ Dashboard

A visual command center for your AI engineering team — track every idea from conception to delivery.

# Option 1: Via CLI (recommended)
uv run aiteam dashboard

# Option 2: Manually via npm
cd dashboard && npm run dev -- --host 0.0.0.0

Open http://localhost:5173 — the dashboard proxies API calls to the backend at localhost:8765.

What makes it powerful

Feature	What you get
🧠 Idea Inbox	Submit ideas in plain language. One click to refine, one click to spin up a full engineering pipeline.
📊 Progress Dashboard	Live percentage counter — pipeline phases × task completion. Watch ideas go from 0% → 100%.
🔀 Agent Workflow Visualization	See exactly which agents work on what, who hands off to whom, and which steps run in parallel — rendered as a color-coded DAG.
✅ Interactive Task Board	Click any task to cycle its status (todo → in progress → done). Each task shows its assigned agent.
🏗️ Pipeline Tracking	6-phase progress bar per pipeline. Advance phases with one click. Auto-completion detection.
👥 Team Roster	See the assembled agent team for each project — who's idle, who's working.
📦 Deliverables	Actual output files attached to each idea. Firmware source code, research reports, pinout diagrams — rendered as styled HTML for comfortable reading.
🌐 Bilingual UI	English / 繁體中文 toggle. 100+ UI strings localized. Preferences saved across sessions.
⏱ Time Tracking	Every task tracks elapsed work time automatically. Pause/resume on blocked/review transitions. See cumulative time per agent across all projects.
🪙 Token Usage	Per-agent token consumption with progress bars. Track LLM costs at a glance — task-level, agent-level, and project-level totals.
🚨 Slow Task Detection	Tasks exceeding time thresholds (configurable, default 2h) auto-trigger the evolution system. Background monitor catches stuck in-progress tasks.
📊 Project Time Analytics	Each project shows total time, agent breakdown with mini-bars, and token usage per agent. Active projects panel on the Dashboard home.
▶️ Auto Schedule	One-click global scheduler in Idea Inbox. Green dot = running, red dot = stopped. Start/Stop directly from the UI — no terminal needed.
👤 Agent Performance Table	Combined time + token view for all agents. Sortable columns with distribution bars. See who's working hardest at a glance.
📈 Daily Token Chart	7-day token usage bar chart on the Dashboard home. Track daily LLM costs and spot usage trends.
🙈 Archive Ideas	Hide unwanted ideas from the inbox without deleting data. Toggle "Show archived" to restore. Soft-delete keeps your database intact.
📋 Per-Project Agent Breakdown	Each project card shows agent-level time and token breakdown. Recently completed projects also visible on the Dashboard home.

⚠️ Important: Execution Flow

The Dashboard is a command center, not an auto-executor. After you start a pipeline, tasks are created in todo state — they won't run by themselves.

Your Idea → Refine → Start Pipeline → Tasks created (all todo)
                                           │
                              ┌────────────┴────────────┐
                              ▼                         ▼
                    ▶ Start Auto Schedule         Manual: click each
                    (Idea Inbox top bar)          task ▶ Run / ⏹ Done
                              │
                              ▼
                    Autonomous mode runs:
                     • Picks next todo task
                     • Sets it to in_progress
                     • Does the work
                     • Sets it to done
                     • Repeats for all projects

To start execution, click "Start Auto Schedule" at the top of the Idea Inbox. It starts directly from the UI — no terminal needed.

The status bar turns green when running, red when stopped. Click Stop to pause. No more wondering why tasks aren't progressing.

Case Studies

Case 1: M5Stack Smart Garden (Embedded Firmware)

"I want a smart garden     →  AI suggests embedded-firmware pipeline
 that auto-waters plants"     6 phases, 8 tasks, 5-agent team

Deliverables:
  📄 firmware.cpp (13 KB)  — Complete ESP32-S3 Arduino firmware
  📄 platformio.ini         — Build configuration  
  📄 pinout.md              — GPIO assignments + calibration table

Case 2: Enterprise AI Transformation Strategy (Business Debate)

"Should we replace 30% of    →  AI suggests quick-prototype pipeline
 staff with AI agents?"          4-phase debate structure

The debate:
  🔴 Automation_Advocate argues for full AI replacement, 30% workforce reduction
  🔵 Collaboration_Advocate argues for human-AI collaboration, zero layoffs  
  ⚖️ CEO rules: "Human-AI Collaboration First" — upskill everyone, automate
     only proven-safe workflows, start with 3 internal pilot projects

Deliverables:
  📄 debate.md (8 KB) — 4-dimension structured debate + 18-month roadmap

🧬 Self-Evolution

The system doesn't just execute — it evolves. Every failure, every research discovery, every agent interaction feeds back into the system, making it smarter over time.

Failure Alchemy

When a task fails or takes too long, the system doesn't just log it. It performs alchemy — turning failure into three kinds of gold:

⏱ Auto-detection: Tasks exceeding their time estimate × 2 (or global 120 min threshold) are automatically reported as timeout failures. The evolution system learns from slow tasks just as it learns from failures.

Output	What it is	How it works
🛡️ Antibody	Prevention strategy	Stored in team memory. Prevents the same class of error from recurring.
💉 Vaccine	Pre-task warning	Injected before similar future tasks. "Before starting I2C driver work, verify timeout is set..."
⚡ Catalyst	Prompt improvement	Injected into agent system prompts. Changes how agents think about the problem class.

Example: An I2C bus hang under high temperature → antibody (add 50ms timeout + bus reset), vaccine (pre-flight checklist for all I2C drivers), catalyst (agent now audits every blocking call for timeout coverage).

Research Loop

Research agents continuously scan for new technology, competitor moves, and emerging patterns:

Research Agent scans  →  Finding submitted  →  Agents debate relevance
                                                    │
                          Accepted findings ─────────┘
                                │
                          Auto-created as Ideas in Inbox
                                │
                          Refine → Pipeline → Tasks → Deliverables

The loop feeds the pipeline. Accepted research becomes real work — without a human typing a single prompt.

Evolution Dashboard

Open http://localhost:5173/evolution to see:

• Health cards: Active antibodies, vaccines, research conversion rate, system health
• Failure alchemy panel: Every failure with its antibody, vaccine, and catalyst
• Research loop panel: All findings with debate notes, acceptance flow, linked ideas

💡 "You leave at night. The system runs. You come back to antibodies that were born while you slept."

🤖 Autonomous Mode

Stop clicking "yes" — the system can run on its own.

Quick Start

From the Dashboard — open Idea Inbox → click "Start Auto Schedule" (top bar).

Or from your terminal:

# Safety: --yes is required to prevent accidental runs
./scripts/autonomous.sh --yes          # Start (every 5 min)
./scripts/autonomous.sh --yes 10m      # Every 10 minutes
./scripts/autonomous.sh --yes 30m      # Every 30 minutes

# Without --yes, the script only prints instructions and exits
./scripts/autonomous.sh                 # ⚠️  Prints help, does NOT start

# Stop the autonomous loop
./scripts/autonomous.sh stop

The Dashboard status bar shows green when running, red when stopped — no need to check the terminal.

What happens when you walk away

1. API server auto-starts if not running
2. Dashboard auto-starts if not running
3. Claude Code enters /loop mode with the autonomous prompt
4. Every cycle, the system:
   • Scans the Idea Inbox for new ideas → refines them → creates pipelines
   • Checks the Task Wall → executes the next high-priority task
   • Advances pipelines when phases are complete
   • Reports failures → antibodies/vaccines/catalysts auto-generated
   • Every 5 cycles: runs research scan → submits findings
5. No permission prompts — .claude/settings.json pre-authorizes common operations

The experience

9:00 PM — You type: ./scripts/autonomous.sh 10m
9:00 PM — You close your laptop and go to sleep

While you sleep:
  9:05 — CEO picks up 2 high-priority tasks
  9:15 — Research agent finds Zephyr 4.0 release notes
  9:25 — Failure alchemy analyzes an I2C timeout bug
  9:35 — 3 tasks marked done, pipeline advances to testing
  ...

7:00 AM — You open the Evolution dashboard
         — 4 antibodies born, 2 research findings accepted
         — All pipelines advanced, tasks completed
         — Zero prompts from you

How it works

Component	Role
scripts/autonomous.sh	Launcher — starts API, dashboard, and Claude Code /loop
scripts/autonomous-prompt.md	The autonomous execution instructions
.claude/settings.json	Permission allowlist — no more yes/no prompts
/loop mode	Claude Code's built-in recurring execution engine

🏗️ Architecture

┌─────────────────────────────────────────┐
│  Dashboard (React 19 + Shadcn UI)       │  Visualization
├─────────────────────────────────────────┤
│  REST API (FastAPI) + CLI (Typer)       │  Control Layer
├─────────────────────────────────────────┤
│  Orchestrator (LangGraph)               │  Orchestration
├─────────────────────────────────────────┤
│  Memory & Knowledge (SQLite)            │  Memory Layer
├─────────────────────────────────────────┤
│  MCP Server (FastMCP) — 107 tools      │  Tool Layer
└─────────────────────────────────────────┘

📦 Supported Hardware

Dev Board	Chip	Status
M5Stack Core S3	ESP32-S3	✅ Deep Integration
M5Stack Core2	ESP32	🔄 Planned
ESP32-DevKit	ESP32	✅ Supported
Raspberry Pi Pico	RP2040	🔄 Planned
STM32F4 Discovery	STM32F407	🔄 Planned
nRF52840 DK	nRF52840	🔄 Planned

🤖 Agent Team

Embedded Engineering (6): firmware-engineer · hardware-engineer · linux-engineer · iot-engineer · sensor-driver-dev · testing-engineer

Software Engineering (5): software-architect · backend-developer · frontend-developer · fullstack-developer · devops-engineer

Management & Quality (5): tech-lead · project-manager · code-reviewer · qa-engineer · technical-writer

Special (2): idea-refiner · rapid-prototyper

📋 Pipeline Templates

Template	Workflow
embedded-firmware	Idea → Pin Planning → HAL → Business Logic → Test → Flash
embedded-linux	Idea → System Design → Driver/App → Cross-compile → Test
web-fullstack	Idea → UI/UX → Frontend → Backend → Deploy
quick-prototype	Idea → MVP → Iterate
research-spike	Technology Research → Feasibility Report

💾 Database Safety

The system auto-protects your data. Every server startup creates a timestamped backup.

# Manual backup
uv run python scripts/db_tool.py backup

# List all backups
uv run python scripts/db_tool.py list

# Interactive restore (with preview)
uv run python scripts/db_tool.py restore

# Check database status
uv run python scripts/db_tool.py status

Backups live in data/db_backups/. Restore previews table counts before overwriting. Never lose data to accidental rm again.

🛠️ Development

# Sync dependencies (including dev tools)
uv sync

# Run tests
uv run pytest

# Lint
uv run ruff check src/ tests/

# Type check
uv run mypy src/ai_embedded_company/

# Start API server
uv run aiteam serve --reload

# Start Dashboard
uv run aiteam dashboard

# Start MCP server
uv run aiteam mcp

📄 License

MIT — Do whatever you want. Build your own AI company.