tools-integration

The built-in <a href="https://claude-me.com/en/glossary/claude-tools/claude-code/" target="_blank" rel="noopener">Claude Code</a> tool used to spawn subagents, renamed from Task in v2.1.63; the old name still works as an alias.

Before every request reaches the model, an organization's own security server inspects the content first and returns an allow-or-deny verdict, which Claude enforces in real time — rather than relying on after-the-fact audits to catch problems retroactively.

An open standard developed by <a href="https://claude-me.com/en/glossary/core-concepts/anthropic/" target="_blank" rel="noopener">Anthropic</a> that lets AI assistants like Claude connect to external data sources and tools (Google Drive, GitHub, databases) using one unified format, instead of requiring a custom integration built from scratch for every new service.

A way for a user to authorize a third-party application without handing over their password directly — instead, the user's existing identity system issues a scope-limited access <a href="https://claude-me.com/en/glossary/core-concepts/token/" target="_blank" rel="noopener">Token</a> that the application uses to act on the user's behalf, strictly within what was authorized.

A bundle that packages tools, commands, and skills into one installable unit, letting Claude gain the full set of capabilities needed for a category of work at once, instead of connecting each piece separately.

A mechanism for marking the repeated, unchanging parts of your prompt (system prompts, long documents, tool definitions) so Claude remembers the processed result — subsequent requests just read from cache instead of reprocessing everything from scratch, at roughly one-tenth the standard input price.

A license category where source code is visible but the reuse and redistribution freedoms typically implied by "open source" aren't guaranteed — frequently mistaken for being the same thing as open source.

A protocol design where every interaction is an independent, self-contained request that doesn't depend on the record of a previous connection, so servers don't need to stay running just to remember connection state — they can deploy directly on spin-up-and-tear-down compute environments.