StackLift
All terms

Glossary

MCP (Model Context Protocol)

An open protocol that lets AI apps connect to tools and data sources through one standard interface.

The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools, data and prompts. An MCP server exposes capabilities in a common format, and any MCP-aware client, such as an assistant or agent, can discover and use them without custom integration code for each pair.

Anthropic introduced MCP in late 2024 and it has since been adopted across the industry, with support from major model providers and developer tools. The specification is now governed in the open, under the Agentic AI Foundation at the Linux Foundation.

Think of it as a common plug. Before MCP, connecting an assistant to a calendar, a database and a ticket system meant three custom integrations per assistant. With MCP, each system offers one server and every compatible client can use it.

MCP standardizes the connection, not the safety. Permissions, approvals and review still need to be designed for each server you allow an agent to use.

Start in 60 seconds

Put this to work on your project.

Get an AI-assisted, architect-reviewed scope, cost range, and timeline. Free, no login.

The complete guide

MCP (Model Context Protocol): the complete guide

Updated November 18, 2026 ยท Published by , Founder, StackLift AI. Written with AI assistance; how we make content.

Short answer

The Model Context Protocol (MCP) is an open standard that lets AI applications connect to tools and data sources in one consistent way. An MCP server exposes tools, resources and prompts, and any MCP-capable client or agent can discover and use them. It reduces custom integration work, but each server still needs careful access control.

What is the Model Context Protocol?

An AI model on its own can only read and write text. To be useful in a business it needs to reach your calendar, your ticket system, your database or your files. Before MCP, every AI application wired up each of those connections its own way, so the same integration was rebuilt many times.

MCP is a shared language for those connections. Anthropic introduced it as an open standard, and it has since been adopted by many AI products, developer tools and platforms. You can think of it as a universal plug: build the connection once as an MCP server, and any application that speaks MCP can use it.

Technically, MCP defines how a client, which is the AI application, talks to a server, which exposes capabilities. Messages use a structured format, and the connection can run locally on your machine or remotely over the network.

Why does MCP matter for a business?

Integration work is a large share of any AI project. Connecting a model to the five systems your team uses means five connectors to build, secure and maintain, and they usually work with only one AI product. A shared protocol changes the math: the connector is written once and reused, and the choice of AI application becomes easier to change later.

  • Lower lock-in: tools keep working if you change the model or the client.
  • Faster pilots: an existing server can be tried in an afternoon.
  • Central control: one server is one place to apply permissions, limits and logs.
  • A path to your own product: customers can bring their AI tools to your data, under your rules.

It is not magic. A badly designed server is still a badly designed integration, and the model still needs good prompts, tests and guardrails.

How does MCP work?

Three roles take part. The host is the application a person uses, such as a chat assistant or a code editor. The client lives inside the host and holds one connection per server. The server is a small program that exposes a system, such as a CRM or a file store.

  1. The host starts or connects to an MCP server.
  2. The client asks the server what it offers: its tools, resources and prompts, each with a description and a schema.
  3. The model sees those descriptions and decides, from the user's request, whether to use one.
  4. If it does, the host asks the user or a policy for approval where needed, then calls the tool on the server.
  5. The server does the work, such as a database query, and returns a structured result.
  6. The model uses the result to continue or to answer.
What a server can expose
CapabilityWhat it isExample
ToolsActions the model may callCreate a ticket, run a query, send a draft
ResourcesData the application can read as contextA file, a record, a schema
PromptsReusable templates a user can pick"Summarize this account"

MCP builds on the idea of tool calling. Tool calling lets a model request an action. MCP standardizes how those tools are described, discovered and connected, so they are portable between applications.

MCP vs. API: what is the difference?

MCP does not replace APIs. Most MCP servers wrap an existing API. The difference is who the interface is for. An API is designed for programmers who read documentation and write code against it. An MCP server describes itself in a form a model can read at run time, so an AI application can discover and use it without custom code for each service.

API and MCP compared
TopicRegular APIMCP server
Main readerA developerAn AI application and its model
DiscoveryRead the docs, write integration codeClient asks the server what it offers
DescriptionReference documentationMachine-readable tool descriptions and schemas
ReuseEach application integrates separatelyOne server works with many clients
Best forSoftware-to-software callsGiving models controlled access to systems

A good design keeps both. The API stays the stable interface for your software, and the MCP server is a thin, well-described layer for AI clients.

How do MCP and AI agents fit together?

An AI agent plans steps and acts toward a goal. To act it needs tools. MCP is a way to give it those tools without hard-coding each one. An agent connected to several MCP servers can search a knowledge base, open a ticket and update a record in one run, choosing the tools it needs.

  • Fewer custom connectors: one protocol instead of one integration per tool.
  • Swappable tools: replace a server without changing the agent.
  • Clear tool descriptions: the model chooses from named, documented actions.
  • Reuse across agents: the same server can serve a chatbot, a coding assistant and a workflow.
  • Better control points: approvals and logging sit at the protocol boundary.

For design guidance see agent orchestration and multi-agent systems. The Model Context Protocol integration page covers how the work is delivered.

What are real-world examples of MCP servers?

Typical servers and what they unlock
Server forWhat the model can doWatch out for
File storageRead and search documentsWhich folders are exposed
DatabaseRun read-only queries to answer questionsWrite access and data volume
Ticketing or help deskLook up, create and update ticketsWho may close or change records
CRMRead account history, log notesCustomer data in prompts
Code repositoryRead code, open branches and pull requestsNever grant direct production access
Internal knowledge baseSearch company policies with sourcesStale or restricted content

Many teams also build a server for their own product, so customers' AI tools can work with it. That is a new kind of interface, in addition to a web app and an API.

What are the security risks of MCP?

Connecting a model to real systems raises the stakes. A tool that can change data can change it wrongly, or be tricked into doing so. The main risks are known and manageable with ordinary security practice.

Risks and controls
RiskWhat can happenControl
Excess permissionA server can do more than the task needsLeast privilege, read-only by default, narrow scopes
Prompt injectionText in a document or page tells the model to misuse a toolTreat tool results as untrusted data, limit tool combinations, require approval for risky actions
Untrusted serversA third-party server behaves badly or changes after you approved itReview code or vendors, pin versions, run in isolation
Credential exposureSecrets stored in configuration leakUse a secrets manager and short-lived tokens
Data leakageSensitive data flows into prompts or logsRedaction, access control per user, careful logging
No audit trailNo one can tell which tool did whatLog every call with user, tool and result

Authentication matters most for remote servers. Use standard sign-in flows, scope access to the signed-in user and never share one broad credential across all users.

When should you build an MCP server?

Build one when more than one AI application needs the same system, when you want customers or partners to connect their AI tools to your product, or when you want one governed place to control what models may do. Skip it when a single feature needs a single simple call, because a direct function call is simpler.

  • More than one AI application needs the same system.
  • Customers or partners want to connect their AI tools to your product.
  • You want one governed place to control and log what models may do.

How do you build and roll out an MCP server?

  1. Choose one system and a short list of tasks users really need.
  2. Define each tool with a clear name, a plain description and a strict input schema.
  3. Start read-only. Add write actions one at a time with approval steps.
  4. Handle authentication and per-user permissions before anything else.
  5. Log every call and set rate limits.
  6. Test with real prompts, including attempts to misuse the tools.
  7. Release to a small group, watch the logs and expand.

Keep tool sets small. A model given dozens of overlapping tools chooses worse than one given a few clear ones. Name tools by the job they do, and write descriptions as if for a new colleague.

What mistakes do teams make with MCP?

  • Exposing a whole API instead of a few well-chosen tools.
  • Giving a server broad write access on day one.
  • Trusting third-party servers without review.
  • Skipping logs, so misuse is invisible.
  • Treating tool output as trusted instructions.
  • Assuming the protocol itself provides security. It describes the connection; you still design access control.

Measure the result by task success, not by number of tools: how often the agent completes the job, how often a person must correct it, how many risky calls were blocked and how long tasks take.

How StackLift approaches MCP, and what it affects in cost and timeline

Effort depends on how many systems you connect, how sensitive the data is, what write actions are allowed and what approvals are required. The protocol is the small part. Access control, testing and monitoring are the larger ones. See AI agent cost for the main drivers.

StackLift treats an MCP server as production software: architect-reviewed scope, least-privilege access, agents that work on feature branches and never touch production secrets, and CI/CD, observability, security review and rollback from sprint one. See AI agent development and run the free estimate.

Questions people ask

Who created MCP?

Anthropic introduced the Model Context Protocol as an open standard. Since then, many AI products and developer tools have adopted it.

Is MCP only for Claude?

No. It is an open protocol, and clients from several vendors support it. A server you build once can work with any MCP-capable client.

Does MCP replace APIs?

No. Most MCP servers sit on top of existing APIs. MCP adds a description of tools that models can read, so AI applications can use the API without custom integration code.

Is MCP secure?

It is as secure as the design around it. The protocol does not remove the need for authentication, least privilege, approvals and logging. Treat servers as production software and review third-party ones before use.

What can an MCP server do?

It can expose tools the model may call, resources the application can read and prompt templates users can pick. Typical servers wrap databases, file stores, ticket systems and code repositories.

Do I need MCP to build an AI agent?

No. Agents can call tools through direct functions or APIs. MCP helps when you want portable, reusable tools across several applications or want to let outside AI tools connect to your product.

Is MCP the same as function calling?

They are related. Function calling is how a model asks for an action inside one application. MCP is a standard way to describe and connect those actions across applications. See tool calling.

Can MCP run locally?

Yes. A server can run on the user's own machine and talk to the application over a local connection, or run remotely over the network. Local servers suit personal tools. Remote servers suit shared systems and need stronger authentication.

Sources

Sources and further reading

The standards, rules, research and official documentation this page draws on.