MCP Clients: Which AI Tools Support MCP

What an MCP Client Does

The MCP client sits between the host application and the MCP servers. It handles the JSON-RPC 2.0 protocol, manages transport connections, performs capability negotiation, and maintains session state. Each client manages a single server connection, and a host typically runs multiple clients simultaneously to connect to multiple servers.

The client's core responsibilities include connection lifecycle management (connecting, initializing, reconnecting, and disconnecting), capability discovery (listing tools, resources, and prompts from the server), request routing (sending tool call requests to the server and receiving responses), and security enforcement (validating that tool calls comply with configured permissions).

From the user's perspective, the client is invisible. Users interact with the host application, which internally creates and manages client connections. The user sees the result of tool calls, not the protocol mechanics that made them possible.

Anthropic Products

Claude Desktop was the first MCP client, shipping with native support when MCP launched in November 2024. It supports local MCP servers via stdio transport, with server configuration through a JSON file. Users can connect multiple servers simultaneously, and Claude can use tools from any connected server within a conversation. Claude Desktop supports all three primitives (tools, resources, prompts) and provides a user consent model where Claude requests permission before invoking tools.

Claude Code, Anthropic's terminal-based coding agent, includes comprehensive MCP support. It supports both stdio and Streamable HTTP transports, can connect to local and remote servers, and integrates MCP tools seamlessly into its coding workflow. Claude Code's MCP integration is particularly powerful because it can combine MCP tools with its native file editing, terminal command, and search capabilities.

The Claude API and Anthropic Agent SDK support MCP through programmatic client implementations. Developers building custom applications with the Claude API can use the MCP TypeScript or Python client SDK to connect their applications to MCP servers, giving their Claude-powered applications access to the same tool ecosystem available in Claude Desktop.

OpenAI Products

OpenAI added MCP support to ChatGPT and the OpenAI platform in early 2025. ChatGPT supports MCP servers for tool access within conversations, extending beyond OpenAI's built-in tools like web browsing and code execution. The integration was a significant moment for MCP adoption, as it confirmed that the protocol was becoming an industry standard rather than an Anthropic-specific technology.

OpenAI subsequently announced the deprecation of its proprietary Assistants API in favor of MCP, with a mid-2026 sunset date. This decision strongly endorsed MCP as the right approach to AI tool integration. Developers who had built custom tools using the Assistants API were encouraged to migrate to MCP server implementations, which would then work across both OpenAI and other platforms.

Google and Microsoft

Google integrated MCP support into Gemini and the Google Cloud AI platform. Gemini users can connect MCP servers to extend the model's capabilities beyond Google's built-in tools. Google Cloud's AI services support MCP for enterprise deployments, allowing organizations to connect their internal tools and data sources to Gemini-powered applications.

Microsoft added MCP support to VS Code through the Copilot extension and to other Microsoft AI products. VS Code's MCP integration is particularly relevant for developers, as it allows Copilot to access project-specific tools, databases, and APIs through MCP servers configured per workspace. Microsoft's Azure AI services also support MCP for cloud-based AI applications.

AI Coding Tools

The AI coding tool ecosystem has been one of the strongest adopters of MCP. Cursor, the AI-native code editor, added MCP support early, allowing developers to connect database servers, documentation servers, and custom tool servers to enhance Cursor's coding assistance. Windsurf (formerly Codeium) similarly supports MCP for extending its AI coding capabilities.

These coding-focused clients tend to have the most mature MCP implementations because their users are developers who benefit directly from tool integrations. A developer using Cursor with an MCP server for their company's API documentation gets more relevant coding suggestions because the model has access to the actual API specifications rather than relying on general training data.

Implementation Completeness Varies

Not all MCP clients implement the full protocol specification. Some support only the tools primitive, ignoring resources and prompts. Some support only stdio transport, not Streamable HTTP. Some implement basic tool calling but lack features like list change notifications, sampling, or progress reporting.

For most users, the tools primitive with stdio transport covers the primary use case: connecting AI models to external functions. Resources and prompts add value for advanced workflows but are not strictly necessary for basic tool integration. When evaluating an MCP client, check which primitives it supports, which transports it handles, and whether it implements features like user consent for tool invocations and error handling for failed tool calls.

Building Custom Clients

Developers building their own AI applications can implement MCP client functionality using the official TypeScript or Python SDKs. The SDKs provide complete client implementations that handle protocol negotiation, transport management, and session lifecycle. Building on these SDKs rather than implementing the protocol from scratch ensures compatibility with the full MCP server ecosystem.

Custom clients are appropriate when building AI applications that need tool access but are not based on one of the major AI platforms. Internal corporate tools, specialized domain applications, and custom agent systems all benefit from MCP client support because it gives them access to the entire ecosystem of MCP servers without building custom integrations for each external service.