Dokumentationerreichbar
Claude Code, Connect Claude Code to tools via MCP
code.claude.com (externe Seite)
Die Herstellerdokumentation zur Anbindung von Tools über MCP in Claude Code. Der Abschnitt zu Tool Search sagt, dass in der Voreinstellung beim Start nur die Namen der Tools und die Anweisungen der Server geladen werden; die vollständige Beschreibung holt das Werkzeug erst, wenn es sie braucht. Abgerufen am 09.10.2026.
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 09.10.2026:
Tool search keeps MCP context usage low by deferring tool definitions until Claude needs them.
bestätigt 10.11.2026Connect a server when you find yourself copying data into chat from another tool, like an issue tracker or a monitoring dashboard.
bestätigt 10.11.2026Once connected, Claude can read and act on that system directly instead of working from what you paste.
bestätigt 10.11.2026HTTP servers are the recommended option for connecting to remote MCP servers.
bestätigt 10.11.2026Stdio servers run as local processes on your machine.
bestätigt 10.11.2026The SSE (Server-Sent Events) transport is deprecated.
bestätigt 10.11.2026claude mcp list shows a health status next to each server it lists
bestätigt 10.11.2026Local scope is the default.
bestätigt 10.11.2026Check .mcp.json into version control so everyone on your team gets the same MCP tools and services.
bestätigt 10.11.2026For security reasons, Claude Code prompts for approval in interactive sessions before using project-scoped servers from .mcp.json files.
bestätigt 10.11.2026Claude Code can’t show that prompt: it loads project-scoped servers without asking.
bestätigt 10.11.2026Authentication tokens are stored securely and refreshed automatically
bestätigt 10.11.2026Servers from claude.ai appear in the list with indicators showing they come from claude.ai.
bestätigt 10.11.2026Connectors from claude.ai are fetched only when your active authentication method is a claude.ai subscription login.
bestätigt 10.11.2026On Team and Enterprise plans, only admins can add servers.
bestätigt 10.11.2026Everything after -- is passed to the server untouched.
bestätigt 10.11.2026Claude Code tries the HTTP transport first and switches to SSE when the server doesn’t accept it.
bestätigt 10.11.2026Use HTTP instead when your server only responds to requests, since HTTP supports OAuth and the claude mcp add --transport flag, while Web
bestätigt 10.11.2026Socket supports neither. Claude Code reads an entry with no type as a stdio server.
bestätigt 10.11.2026This feature only works on mac
bestätigt 10.11.2026OS and Windows Subsystem for Linux (WSL) The entire server entry from that source is used; fields are not merged across scopes.
bestätigt 10.11.2026A server your organization provides through the managedMcpServers managed setting ranks above all of these
bestätigt 10.11.2026A cloned repository can’t approve its own servers
bestätigt 10.11.2026which blocks it in every permission mode
bestätigt 10.11.2026Toggle a server off in the /
bestätigt 10.11.2026mcp panel to stop Claude Code from connecting to it without losing its configuration. The client secret is stored securely in your system keychain (mac
bestätigt 10.11.2026OS) or a credentials file, not in your config This is the supported way to restrict an MCP server to a security-team-approved subset when the upstream authorization server advertises more scopes than you want to grant.
bestätigt 10.11.2026Claude Code runs the command in a shell and gives up on it after 10 seconds
bestätigt 10.11.2026This keeps a project’s .mcp.json or a plugin from sending your Claude Code or cloud provider credentials to a server it names.
bestätigt 10.11.2026Only tool names and server instructions load at session start, so adding more MCP servers has minimal impact on your context window.
bestätigt 10.11.2026Tool search is on by default behind the gateway
bestätigt 10.11.2026Claude Code expects the gateway to forward
bestätigt 10.11.2026Keep descriptions and instructions concise, and put critical details near the start.
bestätigt 10.11.2026since each upfront tool consumes context that would otherwise be available for your conversation
bestätigt 10.11.2026Claude Code displays a warning when MCP tool output exceeds 10,000 tokens and limits output to 25,000 tokens by default.
bestätigt 10.11.2026up to a hard ceiling of 500,000 characters
bestätigt 10.11.2026The idle window defaults to five minutes for HTTP, SSE, Web
bestätigt 10.11.2026Socket, and claude.ai connector servers, and to 30 minutes for stdio servers. An MCP tool call in the main conversation that is still running after two minutes moves to a background task instead of blocking the session.
bestätigt 10.11.2026Claude Code shows that tool’s permission prompt on every call, even in accept
bestätigt 10.11.2026Edits so your own client is responsible for implementing user confirmation for individual tool calls.
bestätigt 10.11.2026This setting uses any-source-true semantics: true in any settings source takes precedence.
bestätigt 10.11.2026Claude Code filters the tool out before Claude sees it, so it never appears in Claude’s tool list.
bestätigt 10.11.2026Anthropic also provides some connectors itself, without you or an admin adding them.
bestätigt 10.11.2026Before v2.1.296, the shorter limit was 2,048 characters, and before v2.1.295 it applied to every tool description, including one loaded through tool search.
bestätigt 10.11.2026All MCP tools loaded upfront, no deferral
bestätigt 10.11.2026Descriptions and instructions from some of the MCP servers that Anthropic hosts at api.anthropic.com reach Claude in full
bestätigt 10.11.2026it doesn’t survive exiting the session
bestätigt 10.11.2026MCP tasks are in public beta in the Claude Code CLI and require Claude Code v2.1.304 or later.
bestätigt 10.11.2026A task on a remote server keeps running after you exit.
bestätigt 10.11.2026Claude Code also skips the prompt in a session you start in bypass
bestätigt 10.11.2026Permissions mode with skipDangerousModePermissionPrompt set in your user settings or in managed settings. On v2.1.304 or later that covers terminal CLI sessions, Agent SDK sessions, and Git
bestätigt 10.11.2026Hub Actions runs. The call comes from the main conversation, not from a subagent
bestätigt 10.11.2026To stop a running task, select it in /
bestätigt 10.11.2026tasks and press x, or ask Claude to. though the server may still finish it
bestätigt 10.11.2026To follow it again, resume the same session.
bestätigt 10.11.2026defer_loading tool fields and tool_reference blocks along with the anthropic-beta header
bestätigt 10.11.2026On v2.1.305 or later, Claude Code can retry a request without tool search after a 400 or 422.
bestätigt 10.11.2026Loaded upfront on models earlier than Opus 4 and Sonnet 4, on the earlier Agent Platform models, and on Microsoft Foundry deployments hosted on Azure
bestätigt 10.11.2026unless CLAUDE_AUTO_BACKGROUND_TASKS is set to 1
bestätigt 10.11.2026In dont
bestätigt 10.11.2026Ask mode, which never prompts, Claude Code denies the call instead. add them to deniedMcpServers by name or by URL pattern
bestätigt 10.11.2026so a server
bestätigt 10.11.2026Url pattern written for the connector’s own URL doesn’t match it no MCP setting or managed-mcp.json reaches them
bestätigt 10.11.2026the ask setting doesn’t reach Claude Code
bestätigt 10.11.2026Claude Code uses that number for both limits.
bestätigt 10.11.2026The session fetches feature flags
bestätigt 10.11.2026Claude keeps working unless you make Claude wait
bestätigt 10.11.2026CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND set to 1
bestätigt 10.11.2026While automatic backgrounding is on, it moves the call after two minutes by default
bestätigt 10.11.2026The moved call is a plain background call rather than an MCP task.
bestätigt 10.11.2026CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND not 1
bestätigt 10.11.2026