An open standard for connecting models to tools and data sources. Before it, every combination of assistant and system needed its own bespoke integration. The protocol defines a common way for a server to expose tools, files and prompts so that any compatible client can use them, which turns a many-to-many problem into a many-to-one.
Technical view
Introduced by Anthropic in late 2024 and since adopted broadly across the industry. Servers expose tools, resources and prompt templates over a defined transport, and clients discover and invoke them on the model’s behalf. The security considerations follow directly from the design: a connected server is untrusted content entering the model’s context, so tool descriptions themselves can carry injected instructions. Permission scoping, human confirmation for consequential actions and careful auditing of third-party servers all matter.
Related
Part of AI Terminology Index, section: Agents and Tools.