MCP works in two directions¶
October 4, 2026 · Kishore Babu Gutta
Scope note: The two directions below describe distinct architectures, not universal MCP support. Released coverage includes initial bounded GitHub MCP task integration; consuming arbitrary provider MCP servers and broad capability search/execute through MCP are not established by this article. Consult the current provider documentation for supported operations and prerequisites.
"MCP support" can describe two different architectures.
They should not be collapsed into one feature.
Direction one: Runtime consumes MCP¶
A provider may expose an MCP server.
Engineering Runtime can eventually treat that as another way to reach the provider:
Intent
↓
Runtime
↓
MCP transport
↓
Provider
In that direction, MCP sits alongside REST, GraphQL, CLI or other provider interfaces.
Runtime still owns admission, policy, context, evidence and the execution contract.
Direction two: Runtime exposes artifacts through MCP¶
There is another possibility:
AI agent
↓
MCP
↓
Runtime capabilities / contracts / operations
↓
Runtime execution
Here MCP is an interface into Engineering Runtime.
An agent can discover governed capabilities or supported operations without becoming the executor itself.
Why the distinction matters¶
The first direction expands how Runtime can reach systems.
The second expands how reasoning systems can discover and invoke Runtime.
Those are related but separate architectural responsibilities.
I call this Two-Directional MCP:
- MCP as an execution transport into providers.
- MCP as an interface exposing governed Runtime artifacts to agents.
Keeping the directions separate makes it easier to preserve the Intent–Execution boundary.