MCP servers

Interpreter can connect to any Model Context Protocol (MCP) server as an extension. This is the main way you extend Interpreter beyond the built-in tools.

What MCP is

Model Context Protocol is an open standard for giving agents tools. An MCP server is a small program that exposes tools, and sometimes resources, that the agent can call. A Slack MCP server exposes tools for reading channels and posting messages. A database MCP server exposes tools for querying tables. A custom MCP server you write can expose anything your business already has an API for.

Each MCP server is a separate process. Interpreter launches it, lists its tools, and routes the agent's tool calls to it.

Add an MCP server

Open Settings, then MCP servers, then Add. You give Interpreter either:

  • the command to launch a local MCP server (for example, an npx or uvx command), or
  • a URL for a remote MCP server.

Save it, and the new server's tools appear to the agent on the next turn.

How calls work

Once a server is connected, the agent sees its tools alongside everything else. Calls are subject to your approvals policy like any other tool. If you have approvals set to ask before sensitive actions, MCP calls follow the same rule.

Good MCP servers to start with

Most teams add MCP servers for systems they already use every day:

  • company chat: Slack, Discord
  • data warehouse: BigQuery, Snowflake, Postgres
  • ticketing: Linear, Jira
  • CRM: HubSpot, Salesforce
  • docs: Notion, Confluence

Most of these have community-maintained MCP servers. Check the MCP ecosystem before writing your own.

Build your own

Any tool you already have an API for can be wrapped as an MCP server. This is the right path when a workflow needs your internal system that isn't on the public list. The MCP spec is small, and there are SDKs for several languages.

Safety

An MCP server can do whatever you allow it to do. Treat a new server the way you would treat installing a new desktop app: start with read-only tools, watch the approvals queue, and widen access only once you trust the workflow.

Give feedback