Line Secretary — AI Secretary Agent for LINE

A LINE Bot that acts as your AI-powered secretary agent. It manages tasks across multiple projects, looks up project details on demand, and holds natural conversations — all through LINE messages.

How It Works

You (LINE message) → "Add a task to web-app: fix the login redirect bug"
Secretary Agent    → "✅ Done! Added 'fix the login redirect bug' to web-app 🔧 Anything else?"

Unlike simple chatbots, Line Secretary is a tool-use agent — it autonomously decides when to create tasks, look up project info, or just chat naturally. Powered by Claude's tool use API.

Architecture

LINE Bot → Webhook (FastAPI) → Claude Agent (Haiku 4.5 + tool use)
                                    ↓
                              Tool execution loop
                              ├─ create_task
                              ├─ list_tasks
                              ├─ complete_task
                              ├─ get_project_info
                              └─ get_pending_summary
                                    ↓
                              Final response → LINE push message

Agent Tools

Tool	Description
list_tasks	List tasks filtered by project/status
create_task	Create a new task in a project
complete_task	Mark a task as done
get_project_info	Read project details (CLAUDE.md)
get_pending_summary	Overview of all pending tasks

The agent runs a loop: it calls tools as needed, processes results, and continues until it has a final text response for the user.

Conversation Examples

Task management — natural language in, action out

You:   "Fix the auth token refresh on mobile-api"
Agent: "✅ Got it! Added 'fix auth token refresh' to mobile-api 🔧 Need anything else?"

You:   "What tasks are still open?"
Agent: "📋 Here's what's pending:
        web-app: 2 tasks
        data-pipeline: 1 task
        Want to see details for any of these?"

You:   "Mark #7 as done"
Agent: "🎉 Nice! Task #7 is done! You're crushing it 💪"

Project lookup — the agent reads your docs on demand

You:   "What does data-pipeline do again?"
Agent: "📖 Here's what data-pipeline is about:
        It handles ETL jobs, syncs data to the warehouse,
        and generates daily analytics reports.
        Built with Python + Airflow 🛠️"

Casual chat — not everything is a task

You:   "Good morning!"
Agent: "Morning! ☀️ Ready to get things done — what's on the agenda today?"

You:   "Ugh, long day..."
Agent: "Hang in there 😮‍💨 You've been putting in solid work. Take a breather!"

Why an Agent, Not Just an API Call?

	Regular Bot	Direct API call	Line Secretary (Agent)
Task management	Keyword matching (/add, /list)	Single prompt → single response	Autonomous — understands intent, picks the right tool
Multi-step actions	Not possible	Manual chaining by developer	Agent loop handles it (e.g., "add task then show summary")
Project awareness	None	Only what's in the prompt	Reads project docs on demand via tools
Conversation	Rigid commands	Stateless (no memory)	Remembers last 10 messages, natural dialogue
Personality	Robotic	Depends on prompt, but dry	Friendly, encouraging, team-member feel
Extensibility	Rewrite code for each feature	Rewrite prompt	Just add a tool definition + handler

The key difference: tool use

A regular API call is one-shot — you send a prompt, get a response, done.

An agent with tool use is a loop — the model can decide to call tools (create a task, look something up), see the results, and then respond. This means it can handle requests like "add a task to web-app and then show me the full pending summary" in a single conversation turn, autonomously chaining multiple actions.

Regular API:    prompt → response (that's it)
Agent:          prompt → [tool call → result → tool call → result → ...] → response

Tech Stack

• Runtime: Python 3 + FastAPI + uvicorn
• Database: SQLite (aiosqlite)
• AI: Anthropic Claude API (Haiku 4.5 with tool use)
• Messaging: LINE Messaging API
• Hosting: Any server (systemd service included)

Setup

Prerequisites

• Python 3.11+
• Anthropic API key
• LINE Developers account

1. Clone & Install

git clone https://github.com/gdalabs/line-secretary.git
cd line-secretary
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Configure Environment

cp .env.example .env
# Edit .env with your keys:
#   LINE_CHANNEL_SECRET=...
#   LINE_CHANNEL_ACCESS_TOKEN=...
#   ANTHROPIC_API_KEY=...

3. Initialize Database

mkdir -p data
python3 -c "import asyncio; from db import init_db; asyncio.run(init_db())"

4. Run

# Development
uvicorn main:app --host 127.0.0.1 --port 3939 --reload

# Production (systemd)
sudo cp secretary-hub.service /etc/systemd/system/
sudo systemctl enable --now secretary-hub

5. Configure LINE Webhook

Expose your server via tunnel (cloudflared, ngrok, etc.) and set the webhook URL in LINE Developers Console.

Customization

Add Projects

Edit config.py to add your own projects:

PROJECTS = {
    "my-project": "Description of your project",
    # ...
}

Add Tools

Add new tools in agent.py:

1. Define the tool schema in TOOLS list
2. Add handler in _execute_tool()

Adjust Personality

Edit the SYSTEM_PROMPT in agent.py to change the agent's personality and behavior.

Cost

Uses Claude Haiku 4.5 — extremely cost-effective:

Usage	Messages/day	Monthly cost
Light	10-20	~$3-5
Normal	30-50	~$5-15
Heavy	100+	~$15-30

License

MIT

Author

GDA Labs