Marcos Taira

Posted on Mar 16

How I connected SAP Business One to WhatsApp using AI agents (open-source)

#agents #ai #automation #opensource

I built an open-source agent runtime that connects AI to SAP B1 via WhatsApp

The problem I was solving

I work as a SAP Business One consultant. Every day I watch managers open the SAP client, navigate through five screens, run a report, export to Excel, and send it on WhatsApp.

The data was always there. The friction was the interface.

So I built Capyra: an autonomous agent runtime that lives in WhatsApp and acts on business systems directly.

What it does

Your team sends a WhatsApp message:

"How many open orders do we have today?"

Capyra queries SAP B1 via Service Layer REST API and replies in seconds:

"7 orders pending approval, total R$ 48,320.
Oldest: order #4821 from Acme Corp, waiting 3 days.
Want me to send a reminder to the approvers?"

No dashboards. No reports. No IT ticket.

Architecture

WhatsApp (Evolution API)
        ↓
Capyra Gateway (WebSocket)
        ↓
Agent Loop (Claude / GPT)
        ↓
Skills (SAP B1, HTTP, Postgres...)
        ↓
Three-layer Memory (pgvector)
        ↓
Immutable Event Store

The agent loop is simple: perceive → plan → act. Each action is recorded as an immutable event. Write operations always require user confirmation.

The Skills system

Skills are directories with two files:

SKILL.md — instructions for the LLM:

# SAP Business One Skill

## Available Tools
### sap_get_orders
Retrieve sales orders with filters...

## Behavior Rules
- Always confirm before write operations
- Format currency as R$ X.XXX,XX

index.ts — tool implementation:

export class SapB1Skill implements SkillExecutor {
  tools = [...]

  async execute(toolName: string, params: Record<string, unknown>) {
    // your SAP B1 Service Layer calls here
  }
}

Any developer can build and share a skill for any system.

Three-layer memory

One of the most interesting parts of Capyra is how it handles memory:

Episodic — last N messages per session. Simple context window.

Semantic — business facts stored as vector embeddings (pgvector). When a message arrives, relevant facts are retrieved by similarity and injected into the prompt.

Procedural — learned preferences per workspace. The agent learns things like "this client always orders in multiples of 50" and stores them as key-value pairs with confidence scores.

Audit trail

Every agent action is an immutable event in PostgreSQL:

SELECT occurred_at, type, tool_name, approved_by
FROM agent_events
ORDER BY occurred_at;

-- message_in        → user sent a message
-- decision          → LLM decided to call a tool
-- tool_call         → tool was called
-- approval_required → user confirmation requested
-- approved          → user confirmed
-- tool_result       → tool returned data
-- message_out       → agent responded

This is critical for enterprise use: you can replay exactly what the agent did and why.

Docker

No Node.js required — just Docker.

curl -O https://raw.githubusercontent.com/marcostaira/capyra/main/docker-compose.hub.yml
curl -O https://raw.githubusercontent.com/marcostaira/capyra/main/.env.example
cp .env.example .env
# edit .env with your values
docker compose -f docker-compose.hub.yml up -d

Images on Docker Hub:

marcostaira/capyra-gateway
marcostaira/capyra-whatsapp

Getting started

Option 1 — CLI (Node.js required):

npm install -g @capyra/cli
capyra init my-company
cd my-company
npm start

Option 2 — Docker only:

curl -O https://raw.githubusercontent.com/marcostaira/capyra/main/docker-compose.hub.yml
cp docker-compose.hub.yml .env.example .env
docker compose -f docker-compose.hub.yml up -d

Both options set up PostgreSQL, Gateway, WhatsApp channel and SAP B1 connector automatically.

What's next

Telegram and Slack channels
Skills marketplace
Agent-to-Agent (A2A) protocol
Computer Use skill for systems without API
Totvs / Omie connectors

DEV Community