What is MCP?
The Model Context Protocol (MCP) is an open JSON‑RPC–based standard that lets a language‑model host (“host process”) connect through lightweight clients to multiple servers that expose prompts, resources and tools in a uniform way — conceptually, “USB‑C for AI.”
Its design borrows from earlier interface standards such as SQL and, more explicitly, the Language Server Protocol (LSP): like LSP decoupled IDEs from language back‑ends, MCP decouples LLMs from data/tool back‑ends, scaling better than bespoke integrations.
Why do we need another protocol?
| Current approach | Key Pain Point(s) |
|---|---|
| Prompt engineering | Manual, fragile, bounded by context window |
| Classic RAG (vector DB + search) | Model cannot execute actions; must paste raw text |
| Function‑calling / plugins | N × M grid of vendor‑specific adapters |
MCP collapses that N × M grid into N + M: each LLM and each integration only implements MCP once.
The N × M Problem in Context
N = number of models or AI platforms (e.g., OpenAI GPT‑4, Anthropic Claude, Mistral, Gemini, etc.)
M = number of tools, APIs, or data sources (e.g., Slack, GitHub, Notion, Salesforce…)
const totalIntegrations = N * M;Example:
| AI Platform | Slack | GitHub | Notion | Salesforce |
|---|---|---|---|---|
| GPT‑4 | ✅ | ✅ | ✅ | ✅ |
| Claude | ❌ | ✅ | ❌ | ❌ |
| Gemini | ✅ | ❌ | ✅ | ❌ |
Each cell represents a custom, handcrafted adapter, often with different logic, formats, and security models. This does not scale.
What MCP Does
MCP proposes a universal adapter model.
Each tool (M) implements the MCP server interface once.
Each model (N) only needs to implement the MCP client interface
That reduces the complexity from N × M integrations to:
const totalIntegrations = N + M;The LLMs don’t care how Slack or GitHub work — they just speak MCP. This means that, for example, Slack’s MCP server doesn’t care if it’s being queried by Claude or GPT‑4. As long as both sides understand MCP, the integration works.
Because servers can stream structured context and expose invocable tools, the model can reason over fresh data, ground responses and trigger side‑effects — without leaking raw credentials into prompts.
It’s something like:
Without MCP: Every phone manufacturer must build a custom charger for every type of outlet.
With MCP: Everyone uses USB‑C. Plug in and go.
Reference Architectures
| Variant | Sketch | When to choose |
|---|---|---|
| A. Central “Context Hub” - Diagram A | Host → single MCP client → hub server → N downstream servers. The hub deduplicates embeddings, enforces org‑wide policy and keeps audit logs. | Enterprises that need a choke‑point for compliance and observability. |
| B. Agent Mesh - Diagram B | Each autonomous agent embeds its own host+client pair; servers run wherever the data lives. Message‑bus (e.g., NATS) coordinates tasks. | Edge/desktop assistants, offline workflows, or environments that favour horizontal scale and fault‑isolation. |
Diagram A:
Key Concepts Mapped
A → B = the LLM process sending JSON-RPC calls through a client.
B → C = the single point of orchestration and governance (hub).
C → D1…D4 = actual tools, databases, services exposed as MCP-compatible endpoints.
Diagram B:
Highlights:
- Each agent is self-contained, able to reason and act.
- The Message Bus allows distributed coordination, pub/sub tasking, and optionally even logging or observability.
- MCP servers live close to their respective data or services (e.g., GitHub, Salesforce, S3).
Both share identical MCP semantics, so teams can migrate from hub to mesh (or hybrid) without rewriting servers.
Domain-Specific Implementations
| Domain | Servers in play | What MCP adds |
|---|---|---|
| Personal AI desktop | filesystem, calendar, email | Fine‑grained permission prompts; no global file‑tree dump in the prompt. |
| Technical copilots (Replit, Sourcegraph Cody) | git, docker, k8s | Secure access to code, logs and CLIs; devs swap Claude↔GPT with no new adapters. Hugging Face Source Swap. |
| Customer support | zendesk, pricing-db, refund-tool | LLM can both answer queries and execute refunds atomically. |
| Strategic analytics | snowflake, newsfeed, slack | Model cross‑joins BI queries with live market news, then pings stakeholders. |
| No‑code workflow automation | Each “Zap” offered as an MCP server; an LLM chains them | Removes brittle REST glue, enabling natural‑language composition. |
Benefits and Drawbacks
Key advantages:
- Portability & vendor‑neutrality – hosts can swap Claude, GPT‑4o or Gemma without touching back‑end code.
- Least‑privilege isolation – each server is sandboxed; the host mediates consent.
- Rapid ecosystem growth – >1 000 community servers in six months.
Open challenges:
- Security – OAuth‑token theft, indirect prompt injection and “tool‑poisoning” remain active research areas.
- Granular auth scopes – today every server invents its own scope vocabulary.
- Extra latency hops – mitigated with keep‑alive streams and server‑side batching.
- Economic incentive – open standards create positive spill‑overs but capture little direct revenue, echoing USB and HTTP history.
Some takeaways for architects & strategists
- Model context becomes an interface, not a hack. Design your data/products to expose MCP servers rather than sprinkling embeddings everywhere.
- Choose topology consciously. Central hubs simplify control; meshes unlock autonomy.
- Budget for security reviews. Token storage, permission scopes and prompt sanitation are the new attack surface.
- Track the standard. Like USB‑C, once an interface tips, it becomes costly to ignore.
MCP is still young, but its blend of open specification, rapid adoption and clear pain‑killer story signals a likely tectonic shift. Experiment in sandbox projects now; by the time your next platform revamp starts, MCP support may be table‑stakes.