Why I built an MCP server to check my docs (and what it taught me) (passo.uno)

The author built a small MCP (Model-Callable Protocol) server—vale-mcp-server—to expose Vale (a deterministic docs linter) as a set of tools LLMs can call directly. MCP is an open protocol that lets you register operations (similar in spirit to OpenAPI) so agents can invoke local or networked code; this server implements three STDIO-type operations: check a document for style, synchronize Vale packages, and verify Vale is installed. Running locally as a command-line service, the server returns tool outputs to LLMs without manual copy‑pasting and can even detect missing dependencies and suggest fixes. This experiment shows why MCP matters: it moves agents from passive text-only helpers to active tool users, enabling pipelines and autonomous workflows—e.g., format conversions, code-snippet validation, metadata retrieval, and “self-healing” docs that read changelogs and update content. The author found a sweet spot combining Vale’s strict rules with LLM nuance for better editorial decisions, but also surfaced practical hurdles: UX friction in running/connecting servers (FastMCP and varied connection options), brittle config steps, and significant security concerns around giving agents read/write access to sensitive data. MCP isn’t plug-and-play yet, but it promises powerful, self-hosted agent tooling for documentation and beyond—if teams invest in safe, well-scoped integrations.