- Type
- Platform, reference implementation
- Stack
- Python, FastAPI, PostgreSQL with pgvector, Redis, Solidity
- Scale
- 18 services, 107 connectors in 11 categories
- Status
- Active development
This is one of Equality Software’s own products, built in-house with our ADLC method. It is not a client engagement.
The problem
AI agents are only useful when they can act on real systems: CRMs, ERPs, messaging tools, data platforms and, increasingly, blockchains. Wiring each agent to each system one by one creates a tangle of integrations with no shared identity, no consistent permissions and no audit trail. That is exactly what regulators and boards worry about.
What we built
AEP-X is a monorepo containing the protocol specifications (eight RFCs), JSON schemas, a microservices core, a connector layer, a Python SDK and a web console. The whole stack starts with one docker compose up.
How governance is built in
- Every connector call is governed. Each connector carries a risk class and a minimum trust score, so an agent with insufficient trust cannot reach a higher-risk system.
- Evidence first. A verification service scores AI outputs before they are acted on, and decisions are written to a tamper-evident audit trail, optionally anchored on-chain.
- Graceful degradation. Blockchain features switch off cleanly when no chain is configured, so the core works without them.
- Traceable scope. Every scope decision traces back to a written ADLC plan and critical evaluation in the repository’s docs.
Engineering practice
The repository includes 40 automated test files, a GitHub Actions workflow and a Jenkins pipeline, plus an operational manual with UML use case and sequence diagrams for the three load-bearing flows. We stood the full stack up against real PostgreSQL with pgvector and Redis, fixed CI failures, and added an SMTP connector and an updated Salesforce adapter.
What it shows for clients
The same patterns apply to any organisation connecting AI agents to its systems: a single governed integration layer, risk-based permissions per connector, verification before action, and an audit trail you can show to a regulator. We bring these patterns to agentic AI projects and AI governance work.