MCP clients compared

Coverage and how to read this comparison

This matrix records behavior in the vendors' public documentation reviewed on 17 August 2026. It is not a promise that every account, plan, operating system, managed-workspace policy, or older installed build exposes the same UI. A dash means the cited documentation does not establish support; it does not prove the capability is impossible. Legacy HTTP+SSE is listed separately from Streamable HTTP because they are different transports.

Client	Local stdio	Remote HTTP	Documented authentication	Documented MCP surface	Approval and trust behavior	Configuration
Claude Desktop	Yes, through local desktop extensions and developer-defined local servers	Yes, through remote web connectors	Extension settings can collect API keys; remote connectors can run their own authorization flow	Official Desktop guidance documents connector-provided actions; this comparison does not infer resources or prompts	Install extensions only from publishers you trust; managed plans can restrict public and custom extensions	Extensions UI; developer configuration for local servers
Claude Code	Yes	Streamable HTTP (http) and legacy SSE	OAuth for HTTP; fixed headers such as bearer/API tokens are supported	Tools, resources, and prompts are documented	Project-scoped .mcp.json servers require approval; choices can be reset	CLI plus project, user, and local scopes
ChatGPT	No direct local server connection	Yes; custom apps connect to remote MCP endpoints	OAuth and the authentication options offered when creating an app; no API-key claim is inferred here	Full MCP apps can expose actions; exact availability depends on plan and rollout	Workspace admins publish apps, can control actions and access, and users may be asked to confirm writes	ChatGPT web settings and workspace app management
Cursor	Yes	Streamable HTTP and legacy SSE	OAuth for remote servers; environment variables can supply API keys to local servers	Tools, prompts, roots, and elicitation are explicitly listed; the cited support table does not list resources	Tool approval is on by default; auto-run is optional	.cursor/mcp.json, global config, one-click installs, or extension API
VS Code	Yes	Streamable HTTP with SSE fallback; explicit sse is also configurable	Request headers/input variables and OAuth with a client ID are documented	Tools, resources, prompts, and MCP Apps	First start normally asks whether the server is trusted; tool calls can require confirmation; local stdio can be sandboxed on macOS/Linux	Workspace or profile mcp.json, guided command, or gallery install
MCP Inspector	Yes	Streamable HTTP and legacy SSE	Custom headers/bearer tokens and interactive OAuth are documented in current Inspector guidance	A testing UI for tools, resources, prompts, protocol messages, and transport behavior	Its local proxy can spawn processes and must not be exposed to untrusted networks	Web UI, TUI/CLI flags, or an MCP configuration file

Important product boundaries

Claude Desktop and Claude Code are different clients. Desktop extensions are designed for locally installed integrations, while Claude Code documents CLI configuration scopes, resource mentions, prompt commands, HTTP headers, and an OAuth login flow. Do not copy a Claude Code command and assume the Desktop UI consumes it.

ChatGPT is remote-first. OpenAI's current guidance says ChatGPT does not directly connect to a server running only on your machine. Custom MCP apps are managed through ChatGPT web and availability varies by plan and workspace controls. Confirm current eligibility before designing distribution around it.

VS Code and Cursor share some configuration ideas, not one guaranteed schema. Both accept JSON-based MCP configuration, but field names, storage locations, trust prompts, input-variable syntax, and fallback behavior are client concerns. Publish a tested example for each client instead of calling one file universal.

Transport and authentication choices

Choose stdio when the server must run on the user's machine or reach local files and tools. The client starts the process and exchanges protocol messages over standard input/output. Distribution then includes runtime installation, environment variables, upgrades, and the risk of executing a package locally.

Choose Streamable HTTP for a shared cloud service. One deployment can serve multiple clients and users, and browser-based OAuth is practical. Legacy SSE remains present in some client configuration and compatibility paths, but new services should target the current Streamable HTTP transport and test the actual clients they intend to support.

API keys are configuration, not interoperability. A local process may receive a secret in its environment; a remote client may attach a header. That does not mean every client exposes a safe way to collect, store, and rotate the same secret. OAuth is a better fit when each user must grant scoped access without giving the MCP client a long-lived upstream credential.

Capabilities are more than tools

Tools are the most consistently visible primitive, but MCP also defines resources and prompts. Client UX differs: Claude Code exposes resource references and prompt commands, VS Code documents resource browsing and slash-command prompts, and Cursor's current protocol table explicitly lists prompts but not resources. Treat a capability as supported only after the client's current official docs or a versioned test establishes it.

Large tool catalogs also have a context cost. A client may load names and schemas into the model's working context, dynamically search them, or let users disable tools. Regardless of client, keep names distinct, descriptions decision-oriented, and schemas compact. Test whether the model selects the intended tool when several tools sound similar.

A practical selection rule

1. Pick the surfaces your users already work in. A technically complete integration that requires switching clients rarely wins.
2. Eliminate clients that cannot reach your deployment model: local stdio versus public or privately tunneled HTTP.
3. Verify the required primitive—tools, resources, prompts, or interactive UI—against current client documentation.
4. Reproduce authentication with the exact client and account tier. Test first authorization, refresh, revocation, and insufficient scope.
5. Review approval controls for writes. A client trust prompt at installation is not the same as per-call confirmation.
6. Publish client-specific configuration and a review date. Avoid a single copy-paste block labeled “works everywhere.”

Next, follow How to connect to a remote MCP server, then apply the MCP server evaluation and trust checklist before granting access.