sea (Skill Engine Agent)

"The ocean accepts all streams."

Imagine an agent runtime as vast and inclusive as the sea—capable of executing any skill you define, turning your intentions into actions.

sea is a lightweight, Go-based agent runtime designed to be the universal engine for AI skills. It combines:
- Skill: Portable capabilities defined in SKILL.md.
- Engine: A robust, single-binary runtime for executing tools and managing state.
- Agent: An intelligent entity that uses the engine to accomplish complex tasks.

Why sea?

Most agent frameworks are heavy, Python-first, or tightly coupled to specific providers. sea takes a different approach:
- 🌊 Engine-First: A stable event protocol + runtime that powers CLI, TUI, or IDE tools.
- 🐚 Skill-First: Capabilities are defined in portable SKILL.md files, not hardcoded logic.
- ⚓️ Safe by Default: Side-effects (file edits, shell commands) are gated behind Human-in-the-Loop (HITL) approvals.
- 🚀 Native Performance: A single, static Go binary. No Python venv hell.

Features

Quick Start

Installation

Option 1: Quick Install (Recommended)

Run this command in your terminal to download and set up sea automatically:

curl -fsSL https://raw.githubusercontent.com/lucas-zan/agent-sea/main/install.sh | bash

Option 2: Manual Download

1. Go to the Releases page.
2. Download the binary for your OS (Windows, macOS, Linux).
3. Rename it to sea (or sea.exe on Windows) and make it executable:
   bash chmod +x sea ./sea help

Option 2: Build from Source

This project requires Go 1.24+.

git clone https://github.com/lucas-zan/agent-sea.git
cd agent-sea
make native
# Output: ./sea

Option 3: Run with Docker

If you don't have Go installed, you can use Docker:

docker compose run --rm dev ./sea chat

2. Configure

Set up your LLM provider (OpenAI-compatible).

cat > .env <<'EOF'
LLM_API_KEY=sk-your-api-key
LLM_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o-mini
LOG_LEVEL=INFO
EOF

3. Run

Start a chat session:

Or run a specific task directly:

```bash
# Ask the agent to do something
./sea chat --initial "Analyze the current directory layout"

# Auto-approve safe actions (optional)
./sea chat --auto-approve

Chat & Sessions

Start a new session:

./sea chat

List previous sessions (history, status, IDs):

./sea chat --list
# Output:
# 📂 Sessions:
#   session_123... - 15 messages - 2026-01-21 10:00

Resume an existing session:

# Append the session ID to resume context
./sea chat session_123...

Using Skills

In sea, capabilities are called "Skills". They are just directories with a SKILL.md file.

Discovery

Skills are loaded from:
1. <project>/.sea/skills/ (Project-specific)
2. ~/.sea/<agent>/skills/ (User-global)
3. ./skills/ (Built-in)

Managing Skills

# List available skills
./sea skills

# Run a skill directly (headless mode)
./sea run novel-init -a name="MyNovel" -a genre="Sci-Fi"

# Validate a skill definition
./sea validate ./skills/my-new-skill

Creating a Skill (SKILL.md)

---
name: my-skill
description: A custom workflow.
allowed-tools: "read_file write_file"
---

# Instructions

1. Read the user's config file.
2. If distinct, update it with new defaults.

Architecture

sea is designed as a modular layered architecture:

graph TD
    User[User] --> CLI[CLI / REPL]
    CLI --> Engine[sea Engine]

    subgraph "sea Runtime"
        Engine --> Runner[Turn Runner]
        Runner --> LLM[LLM Adapter]
        Runner --> Tools[Tool Registry]
        Runner --> Policy[Policy Engine]

        Runner --> MW[Middleware Chain]
        MW --> SkillsMW[Skills]
        MW --> MemoryMW[Memory]
        MW --> PlanMW[Planning]
    end

Commands

Inside the REPL (chat), you can use slash commands:
- /help - List commands
- /init - Initialize persona/config for current dir
- /compress - Compress conversation history (save tokens)
- /quit - Exit

Contributing

sea is an open engine. We welcome contributions!

• Issues: Report bugs or suggest features.
• PRs: Submit improvements to the engine or new built-in skills.

License

MIT