Compare

mpl vs MCP

Short version: MCP is a transport protocol for agent–tool communication. mpl is a contract / quality / audit layer that runs on top of MCP (or A2A, or plain HTTP). They are not substitutes.

The mpl README puts the stack diagram explicitly: mpl sits above MCP, between your agent logic and the transport. The proxy speaks MCP on the wire; the contracts and quality measurement happen in the layer above.

Fair side-by-side

	MCP	mpl
Layer in the stack	Transport & capability protocol (client-server)	Protocol layer above the transport
Primary concern	How an agent discovers and invokes tools, resources, prompts	Typed contracts, quality measurement, audit trails for those invocations
Type system	JSON-Schema tool inputs/outputs, declared per-tool	Versioned `stype` identifiers in a shared registry; same envelope across tools
Quality scoring	Not in scope	QoM profiles: schema fidelity, instruction compliance, groundedness, determinism, ontology adherence, tool outcome
Audit / provenance	Not in scope (clients can log)	BLAKE3 content hash, provenance block, QoM report on every envelope
Deployment shape	Server you run; client library in the agent	Sidecar proxy in front of the MCP server; transparent or strict mode
Cross-namespace policy	Per-tool, declared by the tool server	Glob-pattern policies (e.g. `org.finance.*`) at the proxy boundary
Substitutable?	No — MCP is the transport	No — mpl composes over MCP, doesn’t replace it

When MCP alone is enough

If you’re building a single-agent system, the messages are short-lived, nothing downstream needs forensic auditability, and “the tool either returned a value or raised” is all the validation you need — MCP on its own is fine. Don’t add a layer for a problem you don’t have.

When you also want mpl

• Multiple agents, and you want the same contract enforced regardless of which tool server it lands at.
• A quality score on every call, not just success/failure.
• A tamper-evident record of what happened to point a regulator at.
• Gradual schema rollout — transparent mode first, strict once you’ve seen the traffic.

How they run together

Agent ──(MCP)──▶ mpl proxy ──(MCP)──▶ Your MCP server
                     │
                     ├── validates contract
                     ├── scores QoM profile
                     ├── applies policy
                     └── writes audit record

Sources

mpl README, stack diagram and Phase 1/2/3 scope (github.com/Skelf-Research/mpl).
MCP public spec at modelcontextprotocol.io.

Spot something inaccurate? Open an issue.