The MCP handshake isolates the model from direct server calls, enforcing a controlled tool request flow.
MCP OPEN STANDARD REDEFINES AI TOOL INTEGRATION
*MCP forces AI models to request external tools through an app, not directly. *Runlayer's 2025‑11‑25 spec uses JSON‑RPC envelopes, capability negotiation and built‑in auth to standardize connections.
The Model Context Protocol (MCP) has landed as an open standard that rewrites how AI applications reach for external tools. Built by Runlayer, the protocol forces a strict separation: a server advertises capabilities, an app presents them to the model, and the model requests the server only through the app. No direct server calls, no REST endpoints, no single HTTP tunnel. Instead, MCP wraps JSON‑RPC as a transport envelope, adding method definitions, capability negotiation, transport choices, and authentication. The spec, dated 2025‑11‑25, promises a uniform language for AI agents like Claude, Cursor, or ChatGPT to tap data, prompts, or utilities without bespoke integrations. The protocol does not introduce new wire formats; the client and server still speak plain tools/list and tools/call, preserving existing tool semantics.
How does MCP restructure AI‑tool communication?
MCP forces a three‑party handshake. The server first publishes a capability list—tools, data sources, or prompts it can expose. The AI app reads this list and presents it to the model as a menu of possible actions. When the model selects an action, the app translates the request into a tools/call RPC, sends it to the server, and returns the result to the model. The model never issues a direct HTTP call or raw JSON‑RPC; it only signals intent through the app. This separation isolates the model from network details, enforces authentication, and lets any compliant server plug into any model without code changes.
What tools and transports does MCP support?
MCP does not bind developers to a single transport. It can run over stdin/stdout streams, a single HTTP endpoint, or any custom channel the implementer chooses. The protocol treats JSON‑RPC as the envelope, then adds its own method definitions and capability negotiation steps. Servers expose tools/list and tools/call messages; clients respond with the same verbs. Because the envelope stays JSON‑RPC, existing tooling can be reused. Authentication is baked into the spec, allowing secure handshakes before any tool invocation. The flexibility lets a Runlayer‑hosted server present local file access, external APIs, or prompt libraries to any model that understands the MCP contract.
Why is MCP not just another REST API?
MCP explicitly rejects the REST model. It does not expose a collection of URLs that the model can fetch. Instead, the model never talks to a server directly; it only signals intent through the app. The protocol uses JSON‑RPC as a transport envelope, not the resource‑oriented semantics of REST. This design eliminates the need for the model to manage HTTP verbs, headers, or status codes. It also prevents the model from inadvertently leaking credentials or triggering unintended side effects on a server. By keeping the model insulated, MCP reduces attack surface and aligns with the principle that AI agents should be decision makers, not network clients.
What is the timeline for MCP adoption?
The MCP specification was published on 2025‑11‑25. Runlayer, the operating layer for AI, authored the draft and made it publicly available. No later dates appear in the source, so the protocol is currently at the specification stage. Early adopters can host a small program that exposes tools, data, or prompts and connect it to Claude, Cursor, ChatGPT, or any agent that implements the MCP client side. Because the spec reuses existing JSON‑RPC tooling, integration can begin immediately for developers willing to adopt the new contract. Widespread uptake will depend on community endorsement and the release of client libraries, which the source does not detail.
Frequently asked questions
What is the Model Context Protocol (MCP)?
MCP is an open standard for connecting AI apps to the tools and data they use, where a server advertises capabilities, the app presents them to the model, and the model requests tools through the app.
Who created MCP and when was the spec released?
Runlayer created MCP and the specification is dated 2025‑11‑25.
How does MCP differ from REST APIs?
MCP does not let the model talk directly to a server, uses JSON‑RPC as an envelope, and adds capability negotiation and authentication instead of resource‑oriented HTTP calls.
MCP offers a disciplined bridge between AI models and external capabilities. By mandating a server‑app‑model handshake, it removes the model from raw network exposure and standardizes tool invocation across platforms. The spec’s reliance on familiar JSON‑RPC envelopes lowers the barrier to entry, while its open‑standard status invites any vendor to implement a compliant server. If the AI community coalesces around MCP, the current patchwork of bespoke connectors could give way to a unified, secure ecosystem. Until then, developers must weigh the promise of interoperability against the effort of adopting a brand‑new protocol.
Sources: Hacker News – MCP Explained (https://mcp-explained.com/)