MCP Claude: connect Claude Code to tools safely

MCP Claude quick setup

1. 1Choose one narrow workflow and define the required read or write actions.
2. 2Verify who operates the MCP server and inspect its documentation or code.
3. 3Prefer remote HTTP for a cloud service or stdio for a trusted local process.
4. 4Add the server at local scope for the first private project test.
5. 5Authenticate through the documented OAuth flow when supported.
6. 6Open /mcp and confirm the tools the server exposes.
7. 7Test read-only work before allowing mutations.
8. 8Review logs, output size, failures, revocation, and data handling before rollout.

A safe connection starts with a defined need. Assess the server, connect it with minimal scope, authenticate, examine its tools, test a bounded task, and approve broader use only after evidence.

Use the live Claude Code MCP guide for current commands, transports, scopes, authentication, and administration controls.

What does MCP add to Claude Code?

Without a connector, a developer may copy an issue, log, or database result into a prompt. An MCP server gives Claude Code a structured route to discover and call approved tools. The result can be fresher and more actionable, but the connection also expands access and risk.

Issue tracker

read a ticket and, with approval, update its status.

Monitoring

inspect an incident or error event with bounded access.

Source control

read repository state or prepare a pull-request action.

Database

query a safe dataset through a restricted interface.

Internal API

expose a narrow business operation with validation and audit logs.

Documentation

search maintained product or engineering knowledge.

Do not add a connector just to avoid a short copy-and-paste step. A new server adds software, permissions, credentials, tool descriptions, context use, and failure modes. The workflow should justify that operational cost.

Choose HTTP or stdio transport

Claude Code recommends HTTP for remote cloud services. A local stdio server runs as a process on the user's machine and can suit trusted scripts or tools that need direct local access. The older SSE transport is deprecated in the current guide.

HTTP

remote server reached through a URL, with OAuth or supported headers.

Stdio

local command launched by Claude Code with specified arguments and environment.

SSE

legacy remote transport; use HTTP where the server supports it.

Add a remote HTTP server

Use claude mcp add with the HTTP transport, a short server name, and the official endpoint. Put options before the name. Verify the endpoint with the service owner instead of copying an unverified command from a forum or package description.

1. 1Read the provider's current Claude or MCP setup page.
2. 2Verify the exact HTTPS endpoint and authentication method.
3. 3Add the server at local scope for the pilot.
4. 4Run /mcp inside Claude Code and complete authentication.
5. 5Examine the advertised tools before sending a real request.

Add a local stdio server

For stdio, the double dash separates Claude Code options from the server command and its arguments. Check the executable, package source, install behavior, update path, environment access, filesystem permissions, and network activity before running it.

A local server is code running on your machine. A convenient package command can download and execute changing software. Pin reviewed versions where appropriate and use a controlled install process for team environments.

Pick the smallest MCP configuration scope

Scope controls where the server loads and whether configuration is shared. Start with local scope unless a real workflow requires broader access. Keep credential values out of project configuration and version control.

Local

private to you and the current project. It is the default for a manually added server.

Project

shared with the team through a root .mcp.json file and version control.

User

private to you but available across all projects.

Plugin

bundled with an enabled plugin and managed through that plugin.

Connector

supplied through the Claude product or directory integration.

Managed

deployed and restricted by organization administrators.

When names or endpoints overlap, higher-precedence definitions can win and duplicates can be skipped. Inspect /mcp in the actual project rather than assuming a global configuration is active.

Review project-scoped servers before approval

Project scope stores a shared MCP configuration file in the repository root. Claude Code asks for approval before using project-scoped servers. Review changes to command, arguments, URL, headers, and environment expansion as carefully as executable code.

• The server owner and source are known.
• The transport and endpoint match official documentation.
• Commands and arguments do not include hidden shell behavior.
• Environment references contain names, not committed secret values.
• Requested permissions match the project workflow.
• The team has a removal, revocation, and incident path.

Authenticate remote MCP servers safely

Claude Code supports OAuth for remote HTTP servers that require it. After adding the server, open /mcp and follow the browser flow. Confirm the application name, requested scopes, account, organization, and redirect before approval.

Use the MCP menu to clear authentication when access is no longer needed. Revoke the application at the provider when required. A removed local configuration does not necessarily revoke a remote authorization grant.

The MCP authorization guide explains the OAuth-based flow used to protect sensitive resources and operations.

Apply least privilege to every tool

Grant only the accounts, projects, records, and operations needed for the workflow. Prefer read-only access for exploration. Separate read and write tools when possible, and keep deletion, payment, deployment, user management, and production changes behind explicit approval.

• Use a test workspace or sandbox for initial calls.
• Limit OAuth scopes to the approved subset when supported.
• Use service accounts with narrow roles instead of personal administrator access.
• Keep credentials in an approved secret store or keychain.
• Validate tool input on the server, not only in the model prompt.
• Log actor, tool, target, result, and authorization decision.
• Rate-limit expensive or sensitive operations.
• Make high-impact changes reversible where possible.

Read the official MCP security best practices before exposing sensitive data or write actions.

Inspect tools and context cost

Run /mcp to see connected servers, tool counts, status, and authentication needs. Tool definitions and results consume context. Disable unused servers, return focused results, paginate large datasets, and avoid dumping raw logs or tables into a conversation.

Claude Code warns when a tool result exceeds the documented threshold and has a configurable output limit for tools without their own limit. Raising a limit can increase context use. Fix the server response shape before expanding it.

List, inspect, and remove MCP servers

• List configured servers with claude mcp list.
• Inspect one configuration with claude mcp get followed by its name.
• Remove a server definition with claude mcp remove followed by its name.
• Check status and authentication inside a session with /mcp.
• Reset project-server approvals with claude mcp reset-project-choices.

Troubleshoot an MCP connection

1. 1Confirm the server appears in claude mcp list.
2. 2Inspect its resolved type, URL, command, arguments, and scope.
3. 3Open /mcp and read the exact status or authentication message.
4. 4Check required environment names without printing their secret values.
5. 5Verify network access, endpoint certificates, process paths, and executable permissions.
6. 6Run the server's documented health or local diagnostic step.
7. 7Reduce the test to one read-only tool call.
8. 8Remove or disable the server if behavior remains unexplained.

Authentication and not-found errors require configuration changes and are not automatically retried. A stdio spawn error often means the command is missing from the active PATH. Use a verified absolute executable path when the environment requires it.

Run a safe MCP pilot

1. 1Choose one workflow with a named owner and measurable outcome.
2. 2Review the server, provider, data path, scopes, and retention.
3. 3Connect at local scope to a non-production account.
4. 4Test discovery and one read-only request.
5. 5Test malformed input, denial, timeout, revocation, and large output.
6. 6Add one reversible write action behind explicit approval.
7. 7Document setup, expected tools, logs, support, and removal.
8. 8Move to project or managed scope only after security and user review.

Use Better Design's AI guardrails to define approval, data, audit, and escalation controls for connected agents.

In summary

MCP can replace pasted context with structured, current tool access in Claude Code. That convenience expands the system's authority. Start with one reviewed server, local scope, least privilege, read-only testing, visible approvals, and a documented revocation path.

Frequently asked questions