Context interfaces are, in the words of the memo that introduced the term, descriptions for the LLM of how it can get even more context, should it decide to. Birgitta Böckeler proposed the name in Context Engineering for Coding Agents, saying she could not find an established term for the category, and used it to distinguish this kind of context configuration in coding agents from reusable prompts — instructions and guidance — that are placed in the context directly.
Usage
The memo lists three kinds of context interface: tools, the built-in capabilities of a coding agent such as running bash commands or searching files; MCP servers (Model Context Protocol), custom programs or scripts running locally or on a server that give the agent access to data sources and other actions; and skills (Agent Skills), descriptions of additional resources, instructions, documentation or scripts that the LLM can load on demand when it judges them relevant. It singles out file reading and searching in the workspace as the most basic and powerful context interfaces, since they are how the agent understands the current codebase.
Because each configured interface’s description takes up space in the context, the memo advises thinking strategically about which context interfaces a particular task actually needs. The memo notes that letting the LLM decide when to load context — its example is skills — is what allows agents to run unsupervised, but leaves some uncertainty about whether the context will actually be loaded when expected.