Skip to content
Axiomony documentation

Architecture & security

A shared PostgreSQL authority, individual user experiences, and a clear distinction between what works now and what still needs to be built.

Why the data layer comes first

Traditional software tries to fit many job roles into one predesigned interface. Axiomony separates the shared foundations from individual experiences. It is designed so a business expert can specify the real model and rules, and any employee can ask an AI assistant for the interface that helps them work.

Technical architecture

AI or MCP clientTask or administration request
→
Node.js MCPFixed data tools / internal LLM
→
PostgreSQLData and enforced policies

Data MCP

Fixed, parameterized tools read and write under each user’s PostgreSQL login. The MCP service has no shared superuser credentials. The database enforces GRANT, REVOKE, row-level security, constraints and triggers.

Administration MCP

Authorized administrators submit natural-language instructions. A server-side OpenAI model generates migration SQL based on the real catalog. The server stores a proposed plan in memory, returns the plan for review, and only applies it after explicit confirmation with implement_plan. SQL changes are made as real PostgreSQL objects, not in a parallel middleware rule engine.

Personal applications and hosting

Optional Nginx hosting stores static HTML/CSS/JS under each verified PostgreSQL user. Other users cannot modify those files through the MCP tools. This controls editing only; published pages are served publicly to visitors who can reach Nginx.

The current preferred data route is trusted runtime → Data MCP → PostgreSQL. A standalone static site cannot safely include PostgreSQL passwords. The authenticated server-side runtime needed for independent browser apps has not yet been built.

Security limitations

Not production-ready. The development prototype has not undergone a production security review or end-to-end Docker integration testing. Do not expose its Basic-auth MCP endpoint to the public internet.

AreaCurrent behaviorBefore production
AuthenticationPostgreSQL username/password in HTTP Basic on each MCP requestHTTPS, OAuth/OIDC, secure sessions and brute-force controls
AuthorizationPostgreSQL login; dynamic tool discovery; GRANT/RLS/column rulesReview for privilege escalation and policy gaps
AI-generated SQLStatic checks, plan review and manual confirmationSQL AST validation, cloned-database tests, robust execution sandbox and durable audit
Site hostingNginx serves user-owned static files publiclyDomain/origin isolation, content abuse controls and secure data-backed app runtime
Integration testsNode.js unit tests; no verified live Docker stackAutomated PostgreSQL/Nginx/MCP end-to-end tests
Plan storageShort-lived in-process memory, lost on restartDurable, auditable migration planning and approval

Current status and next priorities

Axiomony v0.5.1 is an early-stage proof of concept. The current stack includes schema introspection, read/write operations, natural-language administration planning, role-aware tools and optional static publishing. A few next steps are especially important:

  1. Introduce public-client-compatible OAuth/OIDC authentication with secure, per-user PostgreSQL identity.
  2. Provide a trusted session-based runtime for user-created browser applications.
  3. Harden database migration execution with policy validation, test databases and durable audit trails.
  4. Run reproducible Docker integration tests across the MCP server, PostgreSQL and Nginx.
  5. Research MariaDB, MySQL Community, Microsoft SQL Server and Firebird as additional database backends. Require database-native read and write authorization rather than middleware-only checks. See the engine comparison and roadmap.

The GitHub repository contains the source code, configuration, unit tests and full README. This website documents the current prototype; it does not host the MCP service itself.