Open Standards & Interoperability
MeetLoyd is built on open standards rather than a proprietary agent format. That matters for one practical reason: the agents you build here have to be able to leave. Everything below is implemented in the platform today.
The standards we speak
| Standard | What it covers | How MeetLoyd uses it |
|---|---|---|
| MCP (Model Context Protocol) | Tools, resources and prompts | Agents consume MCP servers as tools; the platform exposes its own capabilities over MCP. Tracks spec revision 2026-07-28, serving both the current and 2025-era protocol from one endpoint. |
| A2A (Agent-to-Agent) | Agent discovery and task delegation | A published agent card describes what your agents can do, so other systems can find and call them. |
| Agent Skills | Reusable, portable capabilities | SKILL.md files with three-level progressive disclosure, so a skill costs almost nothing until it is actually used. |
| AGENTS.md | Repository instructions for agents | Read when analysing a codebase and written when exporting, so instructions travel with the code. |
| OASF (Open Agentic Schema Framework) | Describing an agent as data | Agents are emitted as OASF 1.1.0 records for registries and directories. |
| OpenTelemetry GenAI | Tracing model calls | Spans carry the gen_ai.* semantic conventions, so agent traffic lands in the observability stack you already run. |
Import and export
Sixteen agent frameworks import and export bidirectionally, including Microsoft Agent Framework, Google ADK, AWS Strands, CrewAI, LangGraph, LlamaIndex, Mastra and Oracle's Agent Spec.
Agents can be brought in from source code or from declarative configuration.
A project whose agents are defined in agents.yaml imports directly, with no
Python to parse — which is how most frameworks are now written.
Frameworks that have been absorbed elsewhere (AutoGen and Semantic Kernel both went into Microsoft Agent Framework) stay importable so you can migrate off them, and are marked deprecated so you are not led onto a dead end.
See MCP Integration and Agent-to-Agent for how each one works in practice.
Commerce protocols
Agent-initiated payments involve several specifications with overlapping names. MeetLoyd implements:
- AP2 (Agent Payments Protocol) — signed mandates describing what an agent is authorised to buy. Mandates are signed with the platform key and can be verified independently against our published JWKS.
- ACP (Agentic Commerce Protocol) — agent checkout against a merchant.
- UCP (Universal Commerce Protocol) — discovery and catalogue, so an agent can find out what is for sale before trying to buy it.
- x402 — per-call stablecoin payment.
The acronym is overloaded across the industry, and the three specifications are unrelated:
- Agentic Commerce Protocol (OpenAI and Stripe) — an agent buying from a merchant. This is the one MeetLoyd implements for checkout.
- Agent Communication Protocol (IBM/BeeAI) — agent-to-agent messaging. MeetLoyd addresses that concern with A2A and SLIM instead.
- Agent Client Protocol (Zed) — an editor talking to a coding agent. Not implemented.
Documentation elsewhere may use any of the three. Check which one is meant.
Why we treat portability as the floor
Connectivity between agents is becoming standard infrastructure, most of it under open foundations. We think that is correct, and we do not try to compete with it — an agent definition you cannot take elsewhere is a liability, not a lock-in advantage.
What MeetLoyd adds sits above the wire: identity minted per agent, policy proven before deployment, enforced budgets, and signed evidence of what happened. Those are the parts that are hard to build and that an auditor asks about. The protocols underneath should be, and are, open.