KSOR
KSOR is a Knowledge System of Record built as a Forward Deployed Engineer product: a company's knowledge lives as governed markdown, and an MCP server lets Claude, ChatGPT and other AI agents answer from that record with citations. When the record does not cover a question, the agent declines instead of guessing.
01 — The Problem
The Problem
Teams deploying AI agents had no single source of truth the agents were allowed to answer from. Knowledge sat in scattered docs with no ownership or approval state, so an agent could repeat a draft, an outdated policy, or simply invent an answer.
02 — The Solution
The Solution
Built KSOR as a repository with two halves: governed markdown under knowledge/ and the serving code under system/. A governance file (.ksor/governance.yaml) defines audiences and approvers, and only approved documents are served. An MCP server acts as the agent door, so Claude, ChatGPT and any MCP-capable client answer from the record with citations and decline when the record is silent. A human docs site renders the same record with a governance ladder, never publishes drafts, and exposes llms.txt for AI crawlers.
03 — Tech Stack
Tech Stack
- MCP
- TypeScript
- JavaScript
- Markdown
- YAML
- llms.txt
04 — Results & Impact
Results & Impact
Gives AI agents one governed, citable source to answer from, with an explicit refusal path instead of hallucinated answers. Released publicly with the source on GitHub.
My role Product design and engineering as a Forward Deployed Engineer product: governance model, MCP server and docs site.