Architecture

Built so partners can trust you with their clients

A simple stack you can run and audit yourself: Python, PostgreSQL and plain JavaScript in Docker. Security rules also live in the database itself.

The problem

What partners worry about

  1. "Will a competitor see my client?"

    Partners often compete for the same clients.

  2. "Where does the AI send my data?"

    Many AI tools send everything to an outside provider.

  3. "What if there is a bug?"

    A single wrong query can expose data in many systems.

The building blocks

  • PostgreSQL with row-level security

    Every query carries the user's role and id. The database itself returns only the rows that user may see. With no context, protected tables return nothing.

  • One instance per company

    Each company gets its own database, secrets and network. Each company's data sits in its own, separate instance.

  • A local AI agent

    The agent talks only to Ollama on the same machine. There is no cloud AI fallback. The AI proposes; only an authenticated click confirms.

  • A read-only gateway

    /mcp for Claude and /api/v1 for scripts. Off by default, read-only, the same role checks on every call, OAuth or API keys.

  • Plugins next to the core

    Plugins react to events or run daily, keep only ids and dates, and read data without writing to CRM tables.

  • An append-only audit log

    Every change is recorded with who, what and when. No role can edit it.

Stack

  • Python 3.12 and Flask
  • Gunicorn
  • PostgreSQL 17
  • Plain JavaScript
  • Docker and Compose
  • Ollama with an open model
  • MCP server
  • OpenAPI

Built in

  • Two-factor sign-in (TOTP) for staff
  • Sign-in attempt limits
  • Session policy and security headers
  • Encrypted backups with a restore drill
  • Vulnerability scanning with pinned scanners
  • Over 4,000 automated tests

Questions

Short answers before you start

Why PostgreSQL row-level security?

Because a bug in one query should not expose another partner's data. The database refuses it even if the code asks.

Can I use another AI model?

Yes. The model is set in the Modelfile; any model Ollama runs can be used.

Where do I read more?

ARCHITECTURE.md, SECURITY.md and the docs folder in the repository.

Ready to start? Run it yourself, or let us roll it out

  • Use it for free

    Download the code from GitHub and run it on your own server. The licence is free for any number of partners.

  • Get help with the roll-out

    We install PartnerType, import your data and fit the system to your sales process.