MCPs

MCPs (Model Context Protocol)

We use MCP and MCP-style integrations so that AI tools and agents can safely call external systems, APIs, and data sources with a consistent, controlled interface.

What we use MCP for

Exposing tools, file systems, and APIs to AI assistants and agents in a standardized way. This reduces ad-hoc integrations and keeps boundaries clear.

Example: tool descriptor

An MCP server exposes tools that the model can call. Conceptually, a tool is described with a name, description, and input schema. Example (shape only; real MCP uses a specific schema):

{
  "name": "read_file",
  "description": "Read the contents of a file under the project root. Use for docs and source code only.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "path": {
        "type": "string",
        "description": "Relative path from repo root, e.g. 'content/docs/ai/index.mdx'"
      }
    },
    "required": ["path"]
  }
}

We keep descriptions precise so the model knows when to use the tool and what to pass. We avoid overly broad tools (e.g. “run any shell command”) in favor of scoped ones (e.g. “read file”, “list directory”).

Security and scope

We limit which tools and resources are exposed. Permissions and audit trails are considered from the start.

• Read-only first — Prefer tools that only read (files, APIs) unless write is required.
• Path and scope limits — Restrict to repo, docs, or a sandbox; no arbitrary filesystem or network.
• Audit — Log tool calls (e.g. which tool, which resource) so we can detect misuse or drift.

Adoption

We prefer MCP-compatible tools and agents where possible so we can compose them into a single ecosystem instead of maintaining one-off integrations.

See also