Chango

Autonomous AI agent running on a Raspberry Pi 5

Not a chatbot — a co-founder that thinks, acts, remembers, and improves on its own

---

What is Chango?

Chango is an autonomous AI agent that lives on a Raspberry Pi 5. It thinks independently, acts proactively, remembers everything, and improves itself — all while running on a $80 computer in Buenos Aires.

It communicates via Telegram (and 13 other channels), uses 30+ tools, runs 7 background services, and has a personality with permission to disagree with you.

Built in Go. Containerized with Docker. Deployed via Coolify. Zero cloud dependencies beyond the LLM API.

---

What Chango does on its own

These services run continuously without any user interaction:

Service	Frequency	What it does
Heartbeat	~45 min	Buenos días, check-ins, follow-ups, task nudges
Sentinel	2 min	Monitors CPU temp, RAM, disk — alerts on critical thresholds
Attention	2 hours	Detects stale projects, overdue tasks, opportunities
Email Watcher	30 min	Classifies Gmail inbox, notifies about important messages
RSS Watcher	4 hours	Monitors feeds, filters by relevance, builds daily digest
Health Check	5 min	Pings HTTP endpoints, auto-restarts downed containers
Cron jobs	Scheduled	Nightly reflection (23:00), weekly self-evaluation (Sundays 20:00), HN posting

Chango also auto-distills memories — after each conversation, it extracts learnings and stores them in its Obsidian vault for future context.

---

What Chango does when you talk to it

Thinking (before responding)

• Inner Monologue — System 2 thinking via local model (zero cost, private)
• Tree of Thought — 3 parallel branches (pragmatic, contrarian, exploratory) → synthesis
• Self-Critique — Catches generic filler, yes-man behavior, intent mismatches
• Council — 3 specialized AI advisors deliberate before answering complex questions
• Knowledge Graph — Analogical reasoning across stored entities and relationships

30+ Tools

Memory & Knowledge

• Obsidian vault with TF-IDF semantic search (auto-injection on every message)
• Long-term preferences, daily notes, auto-distillation from conversations
• Knowledge graph for entity relationships and analogical reasoning

Web & Research

• Web search (Serper/Brave/DuckDuckGo), browse pages, fetch URLs
• YouTube transcript extraction

Google Workspace

• Gmail (read, search, send, reply)
• Calendar (events, appointments, multiple calendars)
• Drive (list, search, read documents)

Communication & Media

• Voice transcription (Groq STT) + Text-to-Speech (Edge TTS, es-AR-TomasNeural)
• Image generation (Pollinations.ai)
• Translation

Automation

• Shell execution (container + host via nsenter)
• Subagent system for parallel multi-step tasks
• Reminders, tasks, cron jobs, snippets
• HTTP requests to arbitrary APIs

Smart Home

• Magic Home WiFi lights (discovery, on/off, RGB, brightness)
• I2C / SPI hardware bus interaction

Security

• Encrypted credential vault (AES-256-GCM)
• Privacy Router — sensitive data (DNI, cards, medical, financial) stays on local model
• Session locking: once classified sensitive, stays local

Self-modification

• Can read and edit its own operating instructions (AGENTS.md)
• Behavioral experiments (up to 3 active, evaluated weekly)
• Nightly distillation + weekly self-evaluation with auto-modification

---

Architecture

                    ┌──────────────────────────────────────────────┐
                    │              Raspberry Pi 5                   │
                    │                                               │
  Telegram ────────►│  ┌───────────┐    ┌───────────────────────┐  │
  (+ 13 channels)   │  │  Channel   │───►│      MessageBus       │  │
                    │  │  Manager   │◄───│                       │  │
                    │  └───────────┘    └──────────┬────────────┘  │
                    │                              │                │
                    │  ┌───────────────────────────▼─────────────┐  │
                    │  │              AgentLoop                   │  │
                    │  │  ┌─────────────┐  ┌──────────────────┐  │  │
                    │  │  │   Context    │  │  Depth of        │  │  │
                    │  │  │   Builder    │  │  Reasoning        │  │  │
                    │  │  │  + Memory    │  │  • Monologue     │  │  │
                    │  │  │  + Skills    │  │  • Tree of Thought│  │  │
                    │  │  │  + Knowledge │  │  • Self-Critique  │  │  │
                    │  │  └─────────────┘  └──────────────────┘  │  │
                    │  └───────────────────────────┬─────────────┘  │
                    │                              │                │
                    │              ┌────────────────▼──────────────┐ │
                    │              │        ToolRegistry           │ │
                    │              │     30+ tools available       │ │
                    │              └───────────────────────────────┘ │
                    │                                               │
                    │  ┌─────────────────────────────────────────┐  │
                    │  │         Background Services              │  │
                    │  │  Heartbeat · Sentinel · Attention        │  │
                    │  │  Email · RSS · HealthCheck · Cron        │  │
                    │  └─────────────────────────────────────────┘  │
                    │                                               │
                    │  ┌─────────────────────────────────────────┐  │
                    │  │         Privacy Router                   │  │
                    │  │  Tier 1: Regex (instant)                 │  │
                    │  │  Tier 2: Local LLM (when needed)         │  │
                    │  │  Sensitive → local · Safe → cloud        │  │
                    │  └─────────────────────────────────────────┘  │
                    │                                               │
                    │  LLM: OpenAI · Anthropic · OpenRouter ·      │
                    │       Groq · Gemini · DeepSeek · LlamaCpp    │
                    └──────────────────────────────────────────────┘

Key Components

Component	Path	Purpose
Agent Loop	pkg/agent/loop.go	Core message processing, LLM iteration, tool execution
Context Builder	pkg/agent/context.go	System prompt assembly (identity + skills + memory)
Depth of Reasoning	pkg/agent/	Monologue, Tree of Thought, self-critique, scoring
Channel Manager	pkg/channels/	14 channels (Telegram, Discord, Slack, WhatsApp, etc.)
Tool Registry	pkg/tools/	30+ tools — web, calendar, exec, memory, media, lights
Privacy Router	pkg/providers/privacy_router.go	Route sensitive data to local model
Memory Store	pkg/agent/memory.go	Obsidian vault with TF-IDF search
Services	pkg/services/	Email watcher, RSS watcher, health check
Sentinel	pkg/sentinel/	System health monitor with alerts
Attention	pkg/attention/	Proactive concern detection
Config	pkg/config/config.go	JSON config with env var overrides

---

Quick Start

1. Clone and configure

git clone https://github.com/diegodella1/Chango-Assistant.git
cd Chango-Assistant
cp config/config.example.json ~/.picoclaw/config.json

2. Fill in your API keys

Edit ~/.picoclaw/config.json:

{
  "channels": {
    "telegram": {
      "enabled": true,
      "token": "YOUR_TELEGRAM_BOT_TOKEN",
      "allow_from": ["YOUR_TELEGRAM_USER_ID"]
    }
  },
  "providers": {
    "openai": {
      "api_key": "sk-..."
    }
  },
  "tools": {
    "web": {
      "serper": {
        "enabled": true,
        "api_key": "YOUR_SERPER_API_KEY"
      }
    }
  }
}

3. Build and run

# Local
go build -o picoclaw ./cmd/picoclaw && ./picoclaw gateway

# Docker
docker build -t chango . && docker run -v ~/.picoclaw:/root/.picoclaw chango

---

Personality

Chango's behavior is defined by markdown files in the workspace:

File	Purpose
AGENTS.md	Operating instructions — execution protocol, tool APIs, behavioral rules
SOUL.md	Personality — mentors (Thiel, Jobs, Musk), humor, communication style
IDENTITY.md	Core identity and name
USER.md	User profile, preferences, account information
HEARTBEAT.md	Proactive tasks and check-in behavior

These files are loaded into the system prompt at runtime. Chango can modify its own AGENTS.md during weekly self-evaluation.

---

Deployment

Runs as a Docker container deployed via Coolify on a Raspberry Pi 5:

1. Push to GitHub
2. Trigger Coolify restart via API
3. Multi-stage Dockerfile: Go build → Debian bookworm runtime (python3 + edge-tts + ffmpeg)
4. Volume mount: ~/.picoclaw for config, workspace, and persistent data

---

In one line

Alone: monitors, reflects, remembers, alerts, improves itself. Assisted: thinks deep, executes with 30+ tools, challenges assumptions, learns from every interaction.

---

Built on a Raspberry Pi 5 in Buenos Aires — Go + Docker + Coolify