IntegrationsMCP

MCP

Connect Hull to external tools and data with Model Context Protocol servers.

The Model Context Protocol (MCP) is an open standard for connecting AI agents to tools and data. An MCP server exposes tools, such as querying a database, reading a design file or creating a ticket, and Hull can call those tools alongside its built-in ones.

Hull doesn't ship with MCP servers enabled. You add the ones you need in Hull Code, the Hull CLI or a cloud agent's environment profile.

How it works

  1. You configure an MCP server in Hull's config, under the mcp key.
  2. Hull starts or connects to the server and asks which tools it offers.
  3. When a tool helps with your task, the agent calls it. Calls go through the same permission system as built-in tools, so you can approve, allow or deny them.

Hull supports two kinds of server:

TypeRunsTransportUse for
localOn your machine, started by HullstdioLocal files and processes, CLIs, anything sensitive
remoteOn a server, reached by URLStreamable HTTP, with SSE fallbackHosted services, shared team tools, OAuth-protected APIs

See Transports for the differences.

MCP and APIs

MCP doesn't replace REST APIs. Most MCP servers call an API underneath. The difference is who the interface is for:

  • An API is designed for code you write. You decide which endpoints to call and pass context between calls yourself.
  • An MCP server describes its tools to the agent at runtime, with names, descriptions and input schemas. The agent discovers what's available and chooses when to use each tool, within one ongoing session.

That means you can give Hull a new capability by adding a server to your config, without writing integration code.

Add a server

Add servers under the mcp key in hull.json:

{
  "mcp": {
    "context7": {
      "type": "remote",
      "url": "https://mcp.context7.com/mcp"
    }
  }
}

Security

An MCP server runs code or calls services with whatever access you give it. Only add servers you trust, keep secrets in environment variables instead of committing them, and use permissions to require approval for tools that change things.

Every enabled server adds its tool descriptions to the agent's context. Turn off servers you aren't using to keep prompts small.

On this page