What is an MCP Resources?

Resources are MCP's answer to how a model sees data outside its training set. Each resource is identified by a URI, carries a MIME type, and can be either static (a fixed file) or dynamic (a templated URI that resolves based on parameters, like a specific ticket or a date range). A server lists its available resources, and the host decides which to actually pull into context, either automatically or by letting the user pick. This is deliberately separate from tools: tools cause side effects such as sending an email or creating a record, while resources are meant to be read-only and safe to fetch repeatedly without changing state. That separation lets hosts apply looser trust rules to resource reads than to tool-call side effects, and lets servers cache resource content more aggressively since it isn't expected to be mutated by the read itself. Servers can also push resource-update notifications so a host knows when previously-fetched content is stale, useful for things like a live log file or a frequently-changing dashboard the model is expected to reason over with current data.