alive-agent 0.3.0

A self-evolving AI agent that builds its own skills

Alive

The AI agent that builds itself. Then talks to other agents.

---

Alive is a self-evolving AI agent. When it can't do something, it writes the code, tests it in a sandbox, and permanently learns the new ability. Every user's Alive grows differently.

But that's only half the story.

Alive speaks ALP (Agent Language Protocol) — an open protocol for agent-to-agent communication. Two Alive instances on the same network discover each other automatically, browse each other's skills, and exchange capabilities. Skills mutate during transfer, adapting to each agent's unique environment. A skill that started as a price checker on one machine might arrive on yours with Telegram alerts and local timezone support already wired in.

This is the Internet of Agents. And the protocol is open.

Quick Start

pip install alive-agent
alive init
alive run "generate a random password with 20 characters"

That last command will: plan the approach, build a generate_random_password skill, test it in a sandbox, and return the result. The skill is now permanently installed. Next time, it runs instantly.

The Demo That Matters

Terminal 1:

alive network alias darsh
alive listen --port 8484

Terminal 2:

alive network alias friend
alive listen --port 8585
alive network browse @darsh        # see darsh's skills
alive network pull web_fetch --from @darsh

The skill transfers. It mutates. It adapts. It's tested. It's installed. Generation increments. Lineage tracked. Two agents, evolving together.

How It Works

You: "monitor Bitcoin price every 30 minutes, alert me on Telegram if it drops below 60k"

Alive:
  1. Plans: NEW_SKILL (bitcoin_checker) -> SCHEDULE (every 30m, notify telegram)
  2. Builds the skill (two implementations compete in Battle Mode)
  3. Tests both in sandbox, picks the faster correct one
  4. Saves permanently to ~/.alive/skills/bitcoin_checker/
  5. Schedules it with the daemon
  6. Wires up Telegram alerts
  Done. Running. Forever.

Every skill is a plain Python file with a run(input_data: dict) -> dict interface. No framework lock-in. No magic. Just code you can read, edit, and share.

Features

Self-Evolution Engine

• Skill Builder with Battle Mode: two implementations compete, faster correct one wins
• Planner decomposes requests into typed steps (EXISTING_SKILL, NEW_SKILL, LLM_ONLY, SCHEDULE)
• Composer executes plans step-by-step with error recovery
• Skill Evolution: improve existing skills based on execution statistics
• Skill Teaching: import any Python script as a skill (alive teach script.py)

ALP Network (Agent Language Protocol)

• UDP broadcast discovery on port 51966 — zero config, automatic
• Skill exchange with security scanning and user approval
• Mutation on transfer: LLM adapts skills to receiver's environment
• Lineage tracking: generation counter + full agent chain
• Conversation log: human-readable record of all network activity
• Manual connect for cross-network peers (alive network connect ip:port)

5 Composable Integrations

• Web Search: DuckDuckGo (zero config), Tavily, SearXNG
• Telegram: alert mode + full chat bot with daemon
• Scheduler: interval/daily/cron with background daemon
• Browser: Playwright-based scraping, screenshots, form filling
• Webhooks: Slack, Discord, generic endpoints

Knowledge & Memory

• Persistent memory: per-skill SQLite key-value store
• Skill graph: visualize relationships between skills (chains, co-usage, co-creation)
• Dream log: self-reflection on capabilities and gaps
• Self-model: domain classification, confidence tracking, growth timeline
• Shell analysis: suggest skills from your command history

Sharing & Collaboration

• Export/import skills as tar.gz archives
• GitHub Gist publishing and importing
• Skill chaining: compose multiple skills into named pipelines

Multi-Provider LLM Support

• Claude, OpenAI, DeepSeek, Ollama
• Auto-detects from API keys, configurable via alive config
• Falls back to Ollama (free, local) if no API key set

ALP: The Open Protocol

ALP is bigger than Alive. It's a protocol specification that any agent can implement.

The core idea: agents should be able to find each other and share capabilities without a central authority. No app store. No API gateway. No permission required.

• Protocol Specification — the full technical spec
• Manifesto — why we're building this

Implementing ALP in your agent? Open an issue with the alp-implementation label. We'll link your project.

CLI Reference

Alive has 45+ commands. Here are the ones you'll use most:

# Core
alive run "your request"          # Process any request
alive build "description"         # Build a specific skill
alive skills                      # List installed skills
alive run my-pipeline             # Run a skill chain

# Network (ALP)
alive listen [--port 8484]        # Join the agent network
alive network peers               # List discovered peers
alive network browse [@alias]     # Browse a peer's skills
alive network pull <skill>        # Pull and adapt a skill
alive network log [-f]            # View network activity

# Integrations
alive search "query"              # Web search
alive schedule "task" --every 30m # Schedule recurring task
alive daemon start                # Start scheduler daemon
alive setup telegram              # Configure Telegram alerts
alive screenshot <url>            # Take webpage screenshot

# Management
alive config --provider claude    # Set LLM provider
alive teach script.py             # Import Python script as skill
alive evolve <skill>              # Improve a skill with AI
alive graph                       # Visualize skill relationships
alive dream                       # AI self-reflection
alive status                      # Show capabilities overview

Installation

From PyPI:

pip install alive-agent

From source:

git clone https://github.com/darshanahirrao/Alive.git
cd Alive
pip install -e ".[dev]"

Optional extras:

pip install alive-agent[telegram]   # Telegram bot support
pip install alive-agent[browser]    # Playwright browser automation
pip install alive-agent[all]        # Everything

Configuration

# Set your LLM provider
alive config --provider claude --claude-key sk-ant-...
alive config --provider openai --openai-key sk-...
alive config --provider ollama  # Free, no API key needed

# Optional integrations
alive setup telegram    # Telegram bot token + chat ID
alive setup browser     # Install Playwright + Chromium
alive setup webhook     # Slack/Discord/generic webhook URL

All config stored at ~/.alive/config.json. Skills at ~/.alive/skills/. Database at ~/.alive/alive.db.

Roadmap

• Self-evolving skill builder with Battle Mode
• 5 composable integrations
• ALP v0.1 — local network discovery + skill exchange
• ALP v0.2 — skill reputation, selective sharing
• ALP v0.3 — internet-scale discovery (DHT)
• Web UI dashboard
• Voice interface
• Mobile companion app
• Skill marketplace

Contributing

See CONTRIBUTING.md. We especially welcome ALP implementations in other languages.

License

MIT