Knowledge
Knowledge HubConcepts and strategic researchGlossaryDefinitions across six domainsFrameworksStructured strategy methodsTopicsConnected subject areasLibrary
PlaybooksPractical strategy guidesBooksBooks and long-form researchOperator VaultMember resources and toolkits
Explore Discover, the Knowledge Hub, playbooks, and the Operator Vault.
Glossary
Model Context Protocol
The Model Context Protocol (MCP) is an open standard for connecting AI applications (hosts) to external context and capabilities through MCP clients and servers that expose resources, prompts, and tools in a consistent way. Specification revision 2026-07-28 centers on a stateless protocol core with self-describing requests; earlier revisions used a stateful initialize handshake and session model. MCP improves interoperability between tools and AI runtimes — it does not by itself replace enterprise identity, authorization, policy, or data-protection controls.
Why it matters
Without a shared protocol, every agent framework reinvents how models discover files, APIs, and prompts. MCP aims to make those integrations portable across hosts so enterprises can expose internal systems once and consume them from multiple AI clients.
The 2026-07-28 specification release (28 July 2026) is the current protocol revision highlighted by the MCP project: a stateless request/response core, per-request protocol information (including protocol version and client capabilities in request metadata), optional discovery instead of a mandatory initialize handshake, revised lifecycle behavior for multi-round interactions, header-based routing for gateways, and further authorization and interoperability hardening. Older clients and servers may still implement the earlier stateful initialization and session model — operators should confirm which protocol version their SDKs speak and plan upgrades under the project’s deprecation windows.
MCP standardizes how context and tools are exposed. Enterprise IAM, least privilege, audit, DLP, and data-residency controls remain the responsibility of the host, identity provider, and surrounding platform — not of MCP alone.
Example (enterprise): A company runs an MCP server for its internal wiki and ticket system. Multiple approved AI hosts can list resources and call tools through MCP, while corporate SSO, scopes, and network policy still decide who may connect and what each tool is allowed to do.
Mechanisms
- Hosts, clients, and servers — the AI application hosts an MCP client that speaks to one or more MCP servers over JSON-RPC.
- Resources, prompts, and tools — standard surfaces for readable context, reusable prompt templates, and invocable actions.
- Stateless protocol core (2026-07-28) — requests are self-describing; any request can land on a server instance without a protocol-level session from the older initialize handshake.
- Optional discovery — clients may call discovery when they want capabilities up front; it is not required for every interaction under the current model.
- Host-side execution and consent — hosts validate and execute tools under their own security and user-consent controls.
Boundaries
- MCP ≠ a model vendor lock-in escape by itself; hosts and servers must still be operated and secured.
- MCP ≠ complete enterprise IAM or authorization — protocol interoperability ≠ access policy.
- MCP ≠ tool calling as a concept; MCP is one standardized way to expose tools and context to hosts that perform tool execution under their own controls.
- Do not describe the older stateful initialization-handshake lifecycle as universally current.
Related topics
None linked yet.
Related concepts & frameworks
Related Discover collections
Further research
No related resources yet.