The Model Context Protocol (MCP) is an open-source standard for connecting AI applications such as Claude to external systems: your data, your tools and your workflows. For a business, it is the difference between pasting a spreadsheet into a chat window and letting Claude read from, and with your approval write to, the systems where the work actually lives, through a defined list of actions.
This guide explains MCP without jargon, how Claude connects to it, the security risks to plan for, and when a plain API is better. Devvista is a software development and growth agency that builds custom MCP servers and integrations for client systems, so we also say when you should not use one.
01 What MCP Is, in Plain Language
The MCP documentation compares the protocol to a USB-C port for AI applications: one standard way to plug an AI application into an external system. Before a standard existed, every pairing of an AI tool and a business system needed its own custom connection. With MCP, a system exposes itself once and any client that speaks the protocol can use it.
The protocol is open, so it is not owned by one vendor's product line. The MCP site lists Claude, ChatGPT, Visual Studio Code and Cursor among the clients that support it. That matters for you because a connection you build is not locked to a single AI product.
MCP only defines how an AI application and an external system talk to each other. In the words of the architecture page, it "does not dictate how AI applications use LLMs or manage the provided context". What the AI is allowed to do, and what you let it see, is decided by how the server and the client are configured. That is where the business risk lives, and most of this article is about it.
02 Hosts, Clients and Servers: Who Does What
MCP has three participants. Knowing them makes every security conversation clearer.
| Participant | What it is | Example |
|---|---|---|
| MCP host | The AI application you use | Claude Code or Claude Desktop |
| MCP client | A component inside the host that keeps one connection to one server | Created by the host for each server you connect |
| MCP server | A program that provides context and actions to clients | A server that exposes your CRM, inbox or accounting data |
A server can run on your own machine, using the stdio transport, or remotely, using Streamable HTTP. A local server typically serves one client, while a remote server typically serves many. Claude Code documents HTTP as the recommended transport for remote servers and stdio for local processes.
03 Tools, Resources and Prompts: What a Server Can Offer
A server exposes three kinds of things, and each has a business meaning.
- Tools are executable functions the AI can call to perform actions, such as running a query, creating a record or sending a request to another system. This is where write access lives.
- Resources are data sources that give the AI context, such as records, files or API responses.
- Prompts are reusable templates that structure how the AI works with the tools, for example a standard way to summarise an account.
The client first asks the server what tools exist, then calls one by name with arguments. The server describes each tool with a name, a description and an input schema, which is how the AI knows what a tool needs. The practical lesson: a tool is a decision about what your business lets an AI do. A server with a tool called "refund_order" and one without it are two very different systems, even if they connect to the same shop.
04 How Claude and Claude Code Connect
In Claude Code, you add a server with a command such as claude mcp add, choosing the transport and, for remote servers, signing in. Many remote servers require sign-in, and Claude Code supports OAuth 2.0 for them (some servers use tokens or API keys instead). A /mcp command inside a session shows the servers and lets you manage them. Servers can be scoped to a single project, to you across projects, or shared with a team through a project file that is checked into the repository.
In the Claude apps, Anthropic calls these connections connectors. Custom connectors use remote MCP servers that Claude reaches from Anthropic's cloud, so the server must be reachable over the public internet unless you allowlist Anthropic's IP ranges. Each connector has per-tool permissions where you can set a tool to Always allow, Needs approval or Blocked, and organisation owners can limit which actions a connected service can take.
The Claude API also has an MCP connector for reaching remote MCP servers from your own application, and the Agent SDK can use MCP servers inside agents you build. The same protocol sits under all of these, which is why one well-built server can serve several tools.
05 What MCP Makes Possible in a Business
Once Claude can use defined tools against your systems, tasks that used to need copy and paste become one request. The patterns we see are these, described as capabilities, not as results.
| Pattern | What the connection allows | Safe starting point |
|---|---|---|
| Look up and summarise | Read customer, order or ticket data and answer a question | Read-only tools, no write access |
| Prepare, then ask | Draft an email, an invoice or a CRM note for a person to approve | A "create draft" tool, never "send" |
| Update with approval | Change a record after a person confirms the exact change | Write tools set to Needs approval |
| Cross-system checks | Compare data across the CRM, inbox and accounting system | Read-only on each system, with a report as output |
| Run an existing workflow | Trigger a workflow you already trust | Expose a scenario as a tool |
The last row already exists in the workflow platforms. Make documents a hosted MCP server that lets AI clients such as Claude run active and on-demand scenarios, and n8n documents both an MCP server trigger and an MCP client node. If you already automate with those tools, MCP can give Claude a controlled way to start what you have already built. Our guide to automating business processes with AI explains how to choose the process first.
We have not opened the documentation for individual CRM, email or accounting vendors' MCP servers, so we do not name them here. Check your vendor's own documentation to see whether an official server exists, and read what it can do before you connect it.
06 The Risks: Why Connecting Claude to a Live System Needs Care
Reading a spreadsheet and changing a customer record are very different powers. Four risks deserve a plan before you connect anything.
- Too much access. A tool or a token that can do everything will eventually do the wrong thing, by mistake or by manipulation. The MCP security guidance recommends a least-privilege approach with minimal initial scopes and warns against wildcard or omnibus scopes such as full-access.
- Prompt injection from outside content. Anything the AI reads can contain instructions: an email, a web page, a support ticket, a PDF. Claude Code's documentation warns: "Servers that fetch external content can expose you to prompt injection risk." If the AI can both read untrusted content and write to your systems, a hostile message could try to steer it.
- Untrusted servers. A server you did not write runs with real access. Anthropic's security page encourages writing your own MCP servers or using servers from providers you trust, and states that Anthropic does not security-audit or manage any MCP server, though it reviews connectors against its listing criteria before adding them to its Directory. The MCP security guidance also describes risks of local servers that run commands on your machine.
- Silent failure. With no logging, you will not know what the AI changed or why.
No safeguard removes these risks completely. Anthropic's own security page says: "no system is completely immune to all attacks." Plan for containment, not perfection.
07 Security Controls That Work: Permissions, Least Privilege, Approvals and Logs
The MCP specification states that there "SHOULD always be a human in the loop with the ability to deny tool invocations". For a business, that translates into controls you can check.
| Control | What it means in practice |
|---|---|
| Least privilege | A dedicated service account for the AI, with only the permissions the task needs. Read-only unless there is a reason to write |
| Narrow tools | Several small, specific tools ("create draft reply") instead of one general "run any query" tool |
| Validation on the server | The specification says servers must validate all tool inputs, implement access controls, rate limit and sanitize outputs. Do not rely on the AI to behave |
| Approvals | Write and delete tools set to require approval. The specification advises clients to show tool inputs before a call so a person sees exactly what will happen |
| Separate reading from acting | Do not give one session both unreviewed external content and powerful write tools |
| Hard blocks outside the prompt | Claude Code documents that CLAUDE.md is context, not enforced configuration, and recommends a PreToolUse hook to block an action regardless of what Claude decides |
| Audit logs | The specification advises clients to log tool usage for audit. Keep logs of who, what, when and the result, on your side as well |
| Credentials | Keep keys in a secrets manager, rotate them, and never paste them into prompts or files in a repository |
In manual mode Claude Code asks before it runs commands or edits files (your default mode may be different, so check), and project-level server definitions need approval before they load in an interactive session. Treat these as one layer, not the only one.
08 MCP or a Plain API Integration?
MCP is not always the right answer. A standard API integration, a webhook or a workflow platform is often simpler, cheaper to run and easier to audit.
| Choose | When |
|---|---|
| Plain API or webhook integration | The steps are fixed and known: when X happens, do Y. No judgement is needed, and the task runs the same way every time |
| Workflow platform (Make, n8n) | You want visible, editable steps and the connectors already exist |
| MCP server | A person or agent needs to ask open-ended questions across a system, or the AI must pick which action to take from a defined list |
| Custom code with the Claude API | One judgement step sits inside a larger product, and you want full control of data flow and errors |
Anthropic's guidance on building agents advises using the simplest solution that works. If your process is "when an order arrives, create an invoice", you do not need an AI choosing tools. Our post on webhooks versus APIs covers the fixed-step case, and business process automation shows what rule-based automation already handles well.
09 What Devvista Builds Around MCP
Devvista builds custom MCP servers and integrations for client systems. As a capability, that means:
- A server that exposes a defined set of read tools and approval-gated write tools for a system you already run, with the access rules written into the server, not only into a prompt.
- Connections between Claude, your workflow platform and your internal systems, using your accounts.
- Logging, alerts and a staging environment so changes are tested before they touch live data.
- Handover: documentation, credentials in your name and a walkthrough for your team.
The work sits within our API development and integration, AI automation and AI agents services, and our workflow automation practice covers the Make and n8n side. For examples of the kind of integration work we do, see our case studies. For instance, one automation build joins Make.com, an AI analysis layer, a CRM and a database with error logging and human checkpoints on high-risk cases (see the credit repair automation case study).
Thinking about connecting Claude to a system you run? Contact us. The form has a budget range field, and we quote after a short call.
10 A Cautious Rollout Plan for Your First Connection
- Choose one system and one question. For example, "find overdue invoices and summarise them".
- Start read-only with a dedicated account that cannot write.
- Use a copy of your data or a test account for the first trial.
- Review the tools list and remove any tool you do not need.
- Add one write tool at a time, set to need approval, and test with real examples.
- Turn on logging and read the first week of logs.
- Review access monthly and disconnect anything you no longer use.
Canada, Pakistan and Saudi Arabia are covered on their own pages: Canada, Pakistan (WhatsApp automation) and Saudi Arabia (chatbots).
? Frequently Asked Questions
→ Related Resources
More on connecting Claude to the systems you run, and the services behind it:
- How to Automate Your Business With Claude: A Step-by-Step Plan
- Claude Code for Business Automation
- The CLAUDE.md File Explained
- AI Agents for Business: What They Can Automate
- Build AI Automation Yourself or Hire a Team?
- What Is AI Automation for Business?
- Webhook vs API: Which Integration Do You Need?
- Business Process Automation Services
- AI & Automation Services
- Workflow Automation Services
- API Development and Integration
- AI Agents
- Case Studies
→ Sources
Opened on 11 November 2026. Documentation changes, so check the current page before you rely on a detail.
- Model Context Protocol: introduction: MCP is an open-source standard for connecting AI applications to external systems.
- Model Context Protocol: architecture: host, client and server; tools, resources and prompts; stdio and Streamable HTTP.
- MCP specification: tools: human in the loop, input validation, access controls, rate limits, confirmation and audit logging.
- MCP security best practices: scope minimization, token passthrough, confused deputy and local server risks.
- Claude Code: connect to MCP servers: transports, scopes, OAuth sign-in and the warning about trusting servers and prompt injection.
- Claude Code security: permission modes, prompt injection safeguards, trusting MCP servers.
- Claude Code memory (CLAUDE.md): CLAUDE.md is context, not enforced configuration; use a hook to block an action.
- Claude Code hooks: a PreToolUse hook can block a tool call.
- Claude Code permissions: which actions ask for approval and how allow and deny rules work.
- Claude connectors (Anthropic Help Center): connectors use remote MCP servers; per-tool Always allow, Needs approval or Blocked.
- Tool use with Claude: client tools run in your application; your code runs the call and returns the result.
- Agent SDK overview: the Agent SDK, the Client SDK and Managed Agents compared.
- Anthropic: Building effective agents: workflows versus agents, start simple, cost and compounding errors, sandbox testing and human checkpoints.
- Make MCP server: AI clients such as Claude can run Make scenarios through Make's hosted MCP server.
- n8n documentation index: lists n8n pages on MCP, human-in-the-loop for AI tools, error workflows and execution logs.