alphatechfinance
IBM on MCP
MCP GUIDE
2026
Two years ago it was one company's side project. Today every major AI platform speaks it, and a Linux Foundation body governs it. This MCP guide explains the Model Context Protocol from first principles — the architecture, the primitives, every specification revision, and the stateless rewrite arriving now.
- What it is: an open standard that turns the M×N integration problem into M+N. Build a server once; any compliant AI client can use it.
- How it works: JSON-RPC 2.0 over two transports (stdio and Streamable HTTP), with three server primitives — tools, resources, prompts.
- Where it stands: the current finalized spec is 2025-11-25. A release candidate labelled 2026-07-28 landed May 21, 2026.
- The big change coming: the RC makes the protocol core stateless — no sessions, no initialize handshake, a new
server/discovercall. - Who owns it: nobody, by design. Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation on December 9, 2025.
Why MCP exists at all
Before MCP, connecting an AI model to a tool meant writing a bespoke integration for that exact pairing. Five models and twenty tools meant a hundred separate integrations to build, test and maintain. Anthropic called it the M×N problem, and it was the single biggest brake on useful AI applications.
MCP collapses that arithmetic. Each model implements the protocol once. Each tool implements it once. You get M+N implementations instead of M×N. It is the same trick USB solved for hardware peripherals — one connector, and everything downstream just works. The comparison to USB-C is now so standard that even the official documentation leads with it.
MCP is not a competitor to function calling. Function calling is the model API — how a model says "I want to call this tool." MCP is the integration layer above it — how that tool gets discovered, described and reached. Most production setups use both.
How MCP is actually wired
Three components, and the relationships between them are strict. Getting this mental model right makes everything else in this MCP guide straightforward.
The host is the AI application — Claude, ChatGPT, VS Code, Cursor. It contains the orchestration logic and, importantly, it owns security: user consent, credential scope and the per-tool allow list all live here, not in the protocol. The client lives inside the host and holds a strict one-to-one relationship with a single server, managing the session, parsing responses and handling errors. The server exposes the actual capability.
Servers offer three primitives and no more. Tools are callable actions, described with a name, a description and a JSON Schema for input. Resources are read-only context identified by URI — files, records, documents. Prompts are reusable templates the server offers the client. Clients can also offer capabilities back: sampling (a server asking the client's model to generate text), elicitation (a server asking the user for more input mid-call) and roots (telling a server which directories it may touch).
Architecture diagrams only take you so far. This walkthrough shows MCP working end to end — a client connecting, discovering tools and calling them — which makes the abstract parts click much faster than reading a spec.
Click to play. Video: YouTube. Open on YouTube
Transports and messaging
Every MCP message is JSON-RPC 2.0 — request, response and notification frames. What changes is how those frames travel.
stdio
The server runs as a local subprocess and exchanges messages over standard input and output. Simplest possible option, and the default for developer tooling on your own machine.
Streamable HTTP
A single endpoint supporting POST and GET, with optional Server-Sent Events for server-to-client streaming. Introduced in 2025-03-26, replacing the original HTTP+SSE transport, which is now deprecated.
Authorization applies to the remote case. Since 2025-03-26 MCP has used an OAuth 2.1-based framework; 2025-06-18 classified servers as OAuth Resource Servers and required Resource Indicators (RFC 8707) so a token issued for one server cannot be replayed against another. The 2025-11-25 revision added OpenID Connect Discovery 1.0 and incremental scope consent via the WWW-Authenticate header.
Every MCP specification revision
One detail confuses almost everyone, so it is worth stating plainly: MCP versions are dates in YYYY-MM-DD form, and that date marks the last backwards-incompatible change — not the announcement day, not a sequential number. That is why the first revision is labelled 2024-11-05 even though Anthropic announced the protocol on November 25, 2024.
The foundation: client-server model, JSON-RPC 2.0 messaging, the tools, resources and prompts primitives, and the stdio and HTTP+SSE transports.
OAuth 2.1 authorization, the Streamable HTTP transport replacing HTTP+SSE, tool annotations, audio content, argument completions and JSON-RPC batching.
Structured tool output, elicitation, resource links in results, servers classified as OAuth Resource Servers, RFC 8707 Resource Indicators, batching removed, and a security best practices page.
The anniversary release: OpenID Connect Discovery, icon metadata, standards-based elicitation enums, tool calling inside sampling, OAuth Client ID Metadata Documents, experimental tasks, and JSON Schema 2020-12 as the default dialect.
Published May 21, 2026 and described as the largest revision since launch. Final publication targeted for July 28, 2026.
The stateless rewrite, explained
If you build on MCP, this is the section that matters most. The 2026-07-28 release candidate does not add features so much as rebuild the foundation, and the driving reason is scale: stateful sessions meant a request had to return to the same server instance, which blocks ordinary horizontal scaling behind a load balancer.
A stateless core
Protocol-level sessions and the Mcp-Session-Id header are removed, so any instance behind normal HTTP infrastructure can answer any request.
No initialize handshake
The initialize exchange disappears. The protocol version and client capabilities ride in each request's metadata instead, with a new server/discover RPC to learn what a server supports up front.
Multi Round-Trip Requests
A single InputRequiredResult pattern replaces server-initiated requests such as roots, sampling and elicitation — one mechanism instead of three.
A real extensions framework
Extensions get reverse-DNS identifiers and independent versioning. Tasks move out of the core into an official extension, and MCP Apps — live since January 26, 2026 — folds in as the first one.
A deprecation policy with teeth
A formal feature lifecycle with a minimum twelve-month window. Under it, Roots, Sampling and Logging are deprecated — not removed — with tool parameters, direct provider APIs and OpenTelemetry as the suggested replacements.
Beta SDKs for the release candidate shipped on June 29, 2026 — Python v2 and TypeScript v2 split into focused packages, plus Go and C#. If you run MCP in production, testing against those betas now is cheaper than migrating under deadline later.
Who owns MCP now
For enterprise buyers this turned out to be the decisive question, and the answer changed on December 9, 2025. Anthropic donated the Model Context Protocol to the newly established Agentic AI Foundation under the Linux Foundation — the same body that stewards Linux, Kubernetes and PyTorch.
The foundation was co-founded by Anthropic, Block and OpenAI, with AWS, Google, Microsoft, Cloudflare and Bloomberg joining as platinum members. Before the donation, a rational procurement objection was: what happens when the vendor who controls this standard changes priorities? Afterwards, that objection largely evaporated. Companies that compete hard at the model layer now collaborate at the infrastructure layer — exactly the pattern that made Linux and Kubernetes durable.
MCP support in Copilot Studio, one of the first major platform adoptions.
Adoption starting with the Agents SDK — the first frontier rival to endorse it.
Gemini models and SDK committed to MCP, with native support at I/O in May.
Bedrock AgentCore reaches general availability with MCP gateway support.
Where to begin with MCP
The honest advice for most people is that you probably do not need to build anything. Thousands of servers already exist, and the fastest path to value is connecting one to a client you already use.
Use before you build
Check the official registry and the community ecosystem first. GitHub, Slack, Postgres and Google Drive servers have existed since launch day.
Target 2025-11-25
Build against the current finalized version, not the release candidate. Then test against the beta SDKs so migration is a checkbox rather than a rewrite.
Treat security as yours
The protocol gives you the surface; the host owns consent, credential scope and allow lists. Assume any tool description reaching your model is untrusted input.
Start with stdio
Local subprocess servers skip the entire authorization layer while you learn the primitives. Move to Streamable HTTP and OAuth when you genuinely need remote access.
If you want to see this kind of tooling applied to real products rather than in the abstract, our ATF live walkthrough shows the workflow end to end, and the Claude Design guide covers a client that leans on exactly these integration patterns.
MCP guide — quick answers
What is MCP?
An open standard that lets AI applications connect to external tools and data through one uniform interface, using JSON-RPC 2.0. Anthropic published the first revision dated 2024-11-05 and open-sourced it on November 25, 2024.
What is the latest specification version?
2025-11-25 is the current finalized version. A release candidate labelled 2026-07-28 was published May 21, 2026, with final publication targeted for July 28, 2026.
Why are versions written as dates?
The date marks the last backwards-incompatible change, not a release number. Backwards-compatible updates do not bump the version. Revisions are marked Draft, Current or Final.
Who controls MCP?
The Agentic AI Foundation under the Linux Foundation, since December 9, 2025. Co-founded by Anthropic, Block and OpenAI, with AWS, Google, Microsoft, Cloudflare and Bloomberg as platinum members.
Is MCP replacing function calling?
No. Function calling is the model API layer; MCP is the integration layer above it. Most production systems use both together.
Scan or download this guide
Scan the code to open the ATF blog on your phone, or download this guide as a single self-contained file to keep or share.
If this guide saved you time, fuel the next one
Alpha Tech Finance is independent. Every guide here is researched from primary sources and given away free — no paywalls, no sponsored takes. A small contribution keeps it that way.
