ZAPLIA
Sign up free

Technical architecture

A CRM that also speaks your AI's protocol.

ZAPLIA isn't a chatbot with a long prompt. It's an agent harness with typed tools, persisted human approval, and a real MCP server — so you can connect your own AI to your data, not just ours.

In plain English

Beyond the AI that works across your connected channels, ZAPLIA lets you connect your own AI tool — like Claude — straight to your business data: your contacts, your catalog, your appointments. That lets you ask your AI to look up or update things for you, your way, without going through support. If none of this means much to you, no worries: the rest of ZAPLIA works just as well without touching any of it. This page is for developers or anyone already using AI tools.

Client MCP

Connect Claude, Cursor, or any MCP client to your account.

MCP (Model Context Protocol) is the open standard for an AI client to use tools from an external server. ZAPLIA exposes a real one, in production, over standard HTTP transport — compatible with Claude Desktop, Claude Code, Cursor, or any generic MCP client.

  1. 1

    Generate your API key

    In Settings → API, create a key with the mcp (or *) scope.

  2. 2

    Connect your client

    Point your MCP client at the server URL with your key as a Bearer token.

  3. 3

    Read and write in your account

    Your AI can now query and operate contacts, products, appointments, and more — the same data you see in your dashboard.

Connection details
URL https://zaplia.app/api/mcp
Header Authorization: Bearer zpk_live_<tu_key>

Every MCP client (Claude Desktop, Claude Code, Cursor…) has its own way to add a remote server — use this URL and this header wherever it asks for the connection.

Quick test with curl

curl -X POST https://zaplia.app/api/mcp \
  -H "Authorization: Bearer zpk_live_<tu_key>" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Your MCP key grants full owner-level access to your account (read and write) — there's no granular per-team-member permission inside MCP yet. Store it with the same care as a password.

21 tools, one engine

The same tools ZAPLIA's sales runtime uses.

No duplicated business logic: every MCP tool calls the same execution engine that connects AI to your business data and supported channels.

Read

  • get_contact_context

    Full history, notes, and sales status for a contact

  • get_contact_by_id

    Look up a contact by ID

  • get_contact_by_channel

    Find a contact by their channel (WhatsApp, email, etc.)

  • get_recent_messages

    Latest messages from a conversation

  • get_knowledge_answer

    Answers with semantic RAG over your Business Brain

  • get_business_context

    Your business profile, catalog, and policies

  • list_metrics

    Sales and activity metrics for your account

  • get_open_tickets

    Open support tickets

  • check_agenda_availability

    Available time slots on your calendar

Write

  • create_contact

    Register a new contact

  • update_contact

    Update a contact's details

  • create_product

    Add a product or service to your catalog

  • set_contact_communication

    Pause or block communication with a contact

  • add_contact_with_welcome

    Register a contact and trigger their welcome message

  • create_follow_up

    Schedule a follow-up reminder

  • create_ticket

    Open a support ticket

  • update_ticket_status

    Update a ticket's status

  • create_order

    Generate a sales order

  • create_appointment

    Book an appointment with your customer

  • book_agenda_slot

    Reserve a specific slot on your calendar

  • send_whatsapp_message

    Send a WhatsApp message to a contact

Why it's a harness, not a prompt on steroids

Four architecture decisions behind that claim.

DAG-based planning

Every task compiles into a dependency graph with an immutable plan hash. Nothing executes blind, and an approval can't be reused if the plan changed after it was issued.

Persisted human approval

The agent's autonomous actions go through a real state machine (pending → approved / rejected / expired), not a "trust me." It auto-expires after 24 hours if no one decides.

Deterministic safety rules

If a draft reply mentions a price, a percentage, or a quote, it always goes to human review — no matter how confident the AI is. The safety net doesn't depend on the model behaving.

Real vector RAG

Your verified facts (Business Brain) are indexed with semantic embeddings, not just plain text — the AI searches by meaning, not exact word matches.

Limits, no fine print

What works today, and what doesn't yet.

  • Your MCP API key grants full owner-level access to your account — there's no granular per-team-member permission inside MCP yet.

  • Actions you run yourself via MCP don't go through the human-approval flow — that flow protects what the autonomous agent decides on its own. If you call the tool from your own client, you're already authorizing yourself.

  • Every account is isolated at the database level (tenant-scoped Row-Level Security) — your key can never see another account's data, no matter what you ask it.

Generate your key and connect your AI in minutes.

The free plan already includes client MCP access.

Create free accountGo to Settings → API