From Knowledge to Context: Building a Google Open Knowledge Format Pipeline That Feeds a Local, Private MCP

The hardest part of putting AI agents to work is not the model — it is the context. The knowledge an agent needs to be useful (what a table means, how two datasets join, which metric is authoritative, what a runbook says to do at 3 a.m., how you make decisions) is scattered across data catalogs, wikis, code comments, tickets, and the heads of a few experienced people. Every AI tool that wants that context re-invents its own way of capturing it, and almost none of it is portable. A useful way to frame the problem is the practitioner’s equation Agent = Model + Harness: an agent’s capability decomposes into the underlying model and the surrounding harness — the orchestration, tools, and context plumbing around it. The model is largely fixed and swappable; the harness is where you actually build. Two open standards, working together, let you build that harness deliberately rather than as bespoke glue. Google Cloud’s Open Knowledge Format (OKF) solves the representation half of the problem. It is a deliberately small, vendor-neutral specification: curated organizational knowledge stored as a directory of Markdown files with YAML frontmatter. No runtime, no SDK, no proprietary catalog. Because it is “just files and just Markdown,” an OKF bundle can be version-controlled alongside code, reviewed in a pull request, rendered on GitHub, published by any static-site generator, and read by any tool or human. The Model Context Protocol (MCP) solves the delivery half. MCP is an open protocol — “a USB-C port for AI applications” — that standardizes how an AI client discovers and pulls context and tools from a server at runtime. If OKF is the library — a curated, portable body of knowledge — then an MCP server is the librarian that answers an agent’s questions from that library on demand.