What MCP is and how it works here
The Model Context Protocol is how agents reach your tools. What it is, how the platform uses it, and what to watch for.
Updated September 26, 2026 · 1 min read
What it is
The Model Context Protocol (MCP) is an open protocol for connecting AI applications to tools and data. An MCP server exposes tools (for example “list open invoices”); an AI application connects to it and lets the model call those tools.
- Many vendors run official MCP servers (Slack, GitHub, Notion, Stripe…), reached over the internet with a token.
- Others run on your own machine (local servers, started as a program), which suits a developer's computer but not a hosted platform.
How the platform uses it
| On the Canvas | At run time |
|---|---|
| Skill: the tool contract (name, when to use it, inputs) | Becomes a tool of the agent it belongs to. |
| Connector: the system and its credential | Points to the tool's MCP server; the credential comes from the Vault. |
| Live run | Anthropic's MCP connector calls the remote MCP server during the agent's turn, with the stored token. |
- Only remote servers (reached over https) run on this platform. A Connector for a local server is left out of runs, with a notice; it still works in the exported Claude Code project, where
.mcp.jsonstarts it on your computer. - For an internal system, host an MCP server yourself and connect it as a Custom MCP server (see Host your own database MCP server).