MCPNews

The Model Context Protocol ecosystem, verified

Current spec 2026-07-28

Claude Code

Spec Support Unverified Confidence: Good (0.85)

Claude Code is an MCP client published by Anthropic; its MCP spec-revision support has not been verified yet. Last verified 2026-08-14.

Does Claude Code support the current MCP specification (2026-07-28)?

Undetermined. We have not verified which MCP specification revision Claude Code supports (as of 2026-08-14).

Which MCP client features does Claude Code support?

Claude Code supports: tools, resources, prompts, roots (last verified 2026-08-14).

When were Claude Code's facts last verified?

2026-08-14. Each fact on this page links to the sources used to verify it.

Claude Code is Anthropic's coding agent for the terminal, IDEs, and CI, and one of the more capable MCP clients on the feature axis: beyond tools it consumes MCP resources and prompts, answers roots/list with the session's working directories (with change notifications), and refreshes a server's capability list live when the server sends list_changed. Servers are added with claude mcp add or configured in .mcp.json and ~/.claude.json across local, project, and user scopes, with project-scoped servers gated behind an approval step.

Transport support is broad. Streamable HTTP is the documented recommendation for remote servers (the config accepts streamable-http as an alias for http), stdio covers local processes, the SSE transport is supported but marked deprecated, and a WebSocket option exists for servers that push events, though without OAuth. Remote HTTP servers authenticate with OAuth using a loopback redirect; Anthropic's connector docs state support for the 2025-03-26, 2025-06-18, and 2025-11-25 authorization specifications with Dynamic Client Registration.

No Anthropic surface names a full MCP spec revision for the client, so spec status is recorded as unknown rather than inferred from the authorization-spec list. Tool results are capped at 25,000 tokens by default, configurable via MAX_MCP_OUTPUT_TOKENS.

How was this verified, and what are the sources?

Verified · method: vendor-docs · confidence 0.85 (Good).

  1. Claude Code MCP documentation - servers configured via claude mcp add, .mcp.json, or ~/.claude.json with local/project/user scopes; "HTTP servers are the recommended option for connecting to remote MCP servers"; the type field "accepts streamable-http as an alias for http"; SSE transport marked deprecated; WebSocket supported for server-push but "HTTP supports OAuth ... while WebSocket supports neither"; answers roots/list with the launch directory plus additional working directories and sends notifications/roots/list_changed (since v2.1.203); supports list_changed refresh of tools, prompts, and resources; servers can declare a claude/channel capability to push events into a session - accessed
  2. Claude connector building docs - protocol features supported across Claude surfaces: tools, prompts, and resources with text, image, and binary content; resource subscriptions and sampling not yet supported; authorization specs 2025-03-26, 2025-06-18, and 2025-11-25 supported with Dynamic Client Registration; Claude Code uses a loopback OAuth redirect; Claude Code max tool result 25,000 tokens (configurable via MAX_MCP_OUTPUT_TOKENS) - accessed