The Model Context Protocol (MCP) is an open protocol for connecting AI applications to external tools and data. It standardises the interface between a client (an AI app or agent runtime) and a server (something exposing tools, resources, or prompts).
The point is decoupling. Before MCP, every integration was bespoke: a tool written for one agent framework had to be rewritten for the next. An MCP server is written once and works with any MCP-compatible client.
The three primitives
| Primitive | What it exposes |
|---|---|
| Tools | Callable functions the model can invoke, with a JSON Schema for arguments |
| Resources | Readable data the client can pull into context |
| Prompts | Reusable prompt templates the server offers |
Agent platforms use tools most heavily — they are what turns a model that can only produce text into one that can act.
Transports
MCP servers are reachable over several transports. Anter connects to remote servers over SSE and Streamable HTTP. The transport affects operational behaviour — connection lifetime, reconnect semantics — but not the tool interface itself.
Anter's two relationships with MCP
Worth separating, because they are easy to conflate:
- Anter as MCP client. You connect MCP servers to your agents, and sub-agents gain access to the server's whole tool catalog at once rather than tools being defined one by one.
- Anter as MCP server. Anter itself exposes an MCP server, so an MCP client such as Claude Code or Claude Desktop can manage your agents conversationally — create them, run them, read back execution traces.
Context injection
A detail that matters for multi-tenant systems: tool arguments can carry per-property session injection, so values like tenant identifiers are filled in at call time and never exposed to the model. The model sees a tool signature without the credential-bearing fields.