glaivio-ai 0.2.1

The opinionated framework for AI-native apps

Glaivio

The opinionated framework for AI-native apps.

Rails did it for web apps. Next.js did it for React. Glaivio does it for AI agents.

from glaivio import Agent, skill

@skill
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"Sunny, 22°C in {city}"

agent = Agent(
    instructions="prompts/system.md",
    skills=[get_weather],
)

agent.run(channel="whatsapp")

That's it. Your agent is live on WhatsApp.

---

How it works

Glaivio gives every AI-native app the same anatomy:

                    ┌──────────────────────┐
                    │   prompts/system.md  │  ← who the agent is
                    └──────────┬───────────┘
                               │
                    ┌──────────▼───────────┐
                    │     🧠  LLM          │  ← the brain
                    │  Claude/GPT/Gemini   │    decides what to do
                    └──────────┬───────────┘
                               │
        ┌──────────────────────┼──────────────────────┐
        │                      │                      │
┌───────▼────────┐  ┌──────────▼──────────┐  ┌───────▼────────┐
│  @skill        │  │  @skill             │  │  @skill        │  ← the arms
│  search_db()   │  │  send_email()       │  │  book_slot()   │    what it can do
└───────┬────────┘  └──────────┬──────────┘  └───────┬────────┘
        │                      │                      │
        └──────────────────────▼──────────────────────┘
                               │
                    ┌──────────▼───────────┐
                    │  📱 WhatsApp/SMS/Web  │  ← the mouth
                    └──────────┬───────────┘    talks to users
                               │
                    ┌──────────▼───────────┐       ┌──────────────────────┐
                    │        User          │       │   👤 Human operator  │
                    │  "that's wrong,      ├──────►│   notified when      │
                    │   I meant X not Y"   │ stuck │   agent is confused  │
                    └──────────┬───────────┘       │                      │
                               │ correction        │  replies "learned:   │
                    ┌──────────▼───────────┐       │   always confirm X"  │
                    │  💡 Self-improvement  │◄──────┘                     │
                    │  agent gets smarter  │                              │
                    │  with every mistake  │                              │
                    └──────────────────────┘

One framework. One way to build. Every AI-native app follows the same pattern.

---

Why Glaivio?

Every developer building an AI agent today faces the same problems:

• Which LLM? How do I swap between them?
• How do I give it memory across conversations?
• How do I connect it to WhatsApp or SMS?
• How do I give it tools without breaking everything?
• How do I deploy it?
• How do I test it when I change the prompt?

There are no standard answers. Every team solves these differently, from scratch, every time.

Langchain and similar SDKs give you the primitives — but you still wire everything together yourself. It's powerful and flexible, but it's not a framework. It's Lego with no instructions.

Glaivio makes the decisions for you.

One way to define skills. One way to add memory. One way to connect channels. One command to deploy. Convention over configuration — the same philosophy that made Rails dominate web development for a decade.

Web era     → Rails      (2004)  — one way to build web apps
Frontend    → Next.js    (2016)  — one way to build React apps
Agent era   → Glaivio    (2026)  — one way to build AI-native apps

If you want full control and flexibility — use Langchain. If you want to ship in hours not weeks — use Glaivio.

---

Install

pip install glaivio-ai

---

Quickstart

Here's a real example — an AI receptionist that books appointments over WhatsApp.

1. Scaffold

glaivio new my-receptionist
cd my-receptionist
cp .env.example .env   # add your ANTHROPIC_API_KEY

2. Write your prompt — prompts/system.md

You are an AI receptionist for Bright Smile Dental.

Your job is to help patients via WhatsApp. Keep replies SHORT — this is a text message.
Max 2 sentences. Never use bullet points or markdown.

Practice info:
- Address: 123 High Street, London
- Hours: Mon-Fri 8am-6pm, Sat 9am-2pm

When booking: ask for name, date and time. Always call check_availability first.
If the slot is taken, offer the alternatives the tool returns.
When rescheduling: cancel the old appointment first, then book the new one.
If medical or urgent, tell them to call the office directly.

3. Generate your skills

glaivio generate skill CheckAvailability
glaivio generate skill BookAppointment
glaivio generate skill CancelAppointment

Fill in the logic — skills are just functions:

# skills/check_availability.py
from glaivio import skill

@skill
def check_availability(date: str, time: str) -> str:
    """Check if a time slot is available. Always call before book_appointment.
    date: YYYY-MM-DD, time: HH:MM 24h format."""
    # call your calendar API here
    return "Available"

# skills/book_appointment.py
from glaivio import skill

@skill
def book_appointment(patient_name: str, patient_phone: str, date: str, time: str) -> str:
    """Book an appointment. Only call after check_availability confirms the slot is free.
    patient_phone: use the current user's ID from context.
    date: YYYY-MM-DD, time: HH:MM 24h format."""
    # call your calendar API here
    return f"Booked {patient_name} on {date} at {time}"

# skills/cancel_appointment.py
from glaivio import skill

@skill
def cancel_appointment(patient_phone: str, date: str) -> str:
    """Cancel an existing appointment on a given date.
    patient_phone: use the current user's ID from context.
    date: YYYY-MM-DD format."""
    # call your calendar API here
    return f"Cancelled appointment on {date}"

4. Wire it up — agent.py

from dotenv import load_dotenv
load_dotenv()

from glaivio import Agent
from skills.check_availability import check_availability
from skills.book_appointment import book_appointment
from skills.cancel_appointment import cancel_appointment

agent = Agent(
    instructions="prompts/system.md",
    skills=[check_availability, book_appointment, cancel_appointment],
    learn_from_feedback=True,
    privacy=True,
)

if __name__ == "__main__":
    agent.run(channel="whatsapp")

5. Run it

glaivio run --channel whatsapp

Point your Twilio webhook at POST https://your-domain/webhook/whatsapp. Done.

---

Core Concepts

Prompts

Write your agent's instructions in plain markdown — no string literals in code:

prompts/
└── system.md

Point your agent at it:

agent = Agent(
    instructions="prompts/system.md",
    ...
)

Glaivio loads it automatically. Edit the prompt without touching agent.py.

---

Skills

Skills are what your agent can do. Define them with @skill:

from glaivio import skill

@skill
def book_appointment(name: str, date: str, time: str) -> str:
    """Book an appointment. date: YYYY-MM-DD, time: HH:MM."""
    # your logic here — call an API, write to a DB, anything
    return "Booked successfully"

The docstring is what the agent reads to decide when to use the skill. Write it clearly.

Skills that need to identify the current user can use user_id — Glaivio injects it automatically into every session:

@skill
def book_appointment(name: str, user_phone: str, date: str, time: str) -> str:
    """Book an appointment. user_phone: use the current user's ID from context."""
    ...

No closures. No wiring. It just works.

---

Agent

from glaivio import Agent

agent = Agent(
    instructions="prompts/system.md",
    skills=[book_appointment, check_availability],
    model="claude-haiku-4-5-20251001",   # or "gpt-4o", "gemini-2.0-flash", "ollama/llama3"
    max_messages=20,                      # context window per session
)

---

Channels

Run your agent on any channel:

agent.run(channel="web")        # browser chat UI + REST API
agent.run(channel="whatsapp")   # Twilio WhatsApp webhook
agent.run(channel="sms")        # Twilio SMS webhook

Or set it in .env:

GLAIVIO_CHANNEL=whatsapp

Then just run:

glaivio run

---

Memory

By default Glaivio uses in-memory storage — zero config, works immediately. Conversation history is lost when the server restarts.

For production, switch to Postgres — history survives restarts, works across multiple instances.

1. Add your database URL to .env:

DATABASE_URL=postgresql://user:pass@localhost/mydb

2. Run migrations once (creates Langgraph checkpoint tables + Glaivio's own session table):

glaivio migrate

Output:

Running Glaivio migrations...
  ✓ Database 'mydb' already exists
  ✓ Langgraph checkpoint tables
  ✓ glaivio_sessions table

✓ Migrations complete.

3. Use PostgresMemory in your agent:

import os
from glaivio import Agent
from glaivio.memory import PostgresMemory

agent = Agent(
    instructions="prompts/system.md",
    memory=PostgresMemory(url=os.getenv("DATABASE_URL")),
)

Glaivio creates and maintains two sets of tables:

• Langgraph checkpoint tables — full message history per user, keyed by their ID
• glaivio_sessions — your own table tracking user_id, channel, message_count, tokens_used, last_seen

---

Knowledge

Drop files in and the agent searches them automatically:

from glaivio.knowledge import Knowledge

agent = Agent(
    instructions="prompts/system.md",
    knowledge=Knowledge(["./faqs.md", "./pricing.pdf", "./policies.txt"]),
)

Supports .txt, .md, .pdf. No configuration needed.

pip install glaivio-ai[knowledge]

---

Human Handoff

When the agent can't handle something, escalate to a human:

from glaivio.handoff import handoff_to_human

agent = Agent(
    instructions="prompts/system.md",
    on_confusion=handoff_to_human(notify="whatsapp:[phone]"),
)

The agent detects confusion, notifies your team via WhatsApp/SMS, and holds the conversation until a human takes over.

---

Privacy

Automatically redact PII before it reaches the LLM:

agent = Agent(
    instructions="prompts/system.md",
    privacy=True,  # redacts phone numbers, emails, NHS numbers, NI numbers
)

---

Learning from Feedback

The agent learns from user corrections automatically:

agent = Agent(
    instructions="prompts/system.md",
    skills=[book_appointment],
    learn_from_feedback=True,
)

When a user says "that's wrong, I said Tuesday not Wednesday" — the agent extracts the correction, stores it, and applies it to all future conversations:

[Learned from past conversations]
- Always book the exact day the user specifies, never the next day
- When user says Tuesday, confirm Tuesday before booking

Corrections persist in .glaivio/corrections.json. The agent gets smarter over time without any manual prompt editing.

---

Structured Extraction

Extract structured data from natural language — no prompt writing:

from pydantic import BaseModel
from glaivio import extract

class BookingRequest(BaseModel):
    name: str
    date: str   # YYYY-MM-DD
    time: str   # HH:MM
    reason: str = "Appointment"

booking = extract(BookingRequest, from_message="I need Tuesday 10am, I'm John Smith")
# → BookingRequest(name="John Smith", date="2026-03-25", time="10:00", reason="Appointment")

---

CLI

glaivio new my-app                      # scaffold a project
glaivio run                             # start the agent
glaivio run --channel whatsapp          # start on a specific channel
glaivio generate skill BookAppointment  # generate a skill stub
glaivio migrate                         # run database migrations (Postgres only)
glaivio test                            # run evaluations
glaivio deploy                          # generate Railway deployment files
glaivio deploy --target render          # generate Render deployment files
glaivio deploy --target fly             # generate Fly.io deployment files

---

Evaluations

Test your agent like you test your code:

# tests/test_booking.py
from glaivio.testing import eval, EvalCase

@eval
def test_booking(agent):
    return [
        EvalCase("I want Tuesday 10am", "booked", "basic booking"),
        EvalCase("Cancel my appointment", "cancelled", "cancellation"),
        EvalCase("Do you accept BUPA?", "bupa", "insurance FAQ"),
    ]

glaivio test
# → 3/3 passed ✅

Change your instructions and run again — regressions are caught automatically.

---

Supported Models

Prefix	Provider	Example
claude-	Anthropic	claude-haiku-4-5-20251001
gpt-	OpenAI	gpt-4o
gemini-	Google	gemini-2.0-flash
ollama/	Local (Ollama)	ollama/llama3

---

Deploy

glaivio deploy

Generates a Dockerfile, docker-compose.yml, and railway.toml.

railway login
railway up

Done. Your agent is live.

---

License

MIT