MCP / 2025-09-12
Ziksaka: A Knowledge Graph MCP Server
An MCP server that turns a personal Obsidian vault into a typed knowledge graph agents can query in one call.
Role
Architect and builder. Graph schema, server, deployment.
Outcome
Agent context prep dropped from ~15 min of manual gathering to a single tool call. 1,400+ notes and 3,000+ typed edges queryable through one MCP interface.
Most personal knowledge tools are built for humans to read. This one is built for agents to query.
The problem: my agents had no single place to trust for context. Who a stakeholder is, what happened in the last meeting, what a project connects to. All of it lived as flat markdown. Readable by me, useless to an agent that needs a straight answer before it acts.
The reframe: stop treating the vault as notes. Treat it as a typed graph. Entities, typed edges, timelines. Then expose it through MCP tools like resolve_entity, get_dossier, and timeline, so an agent asks a question and gets structured truth back, not a pile of files to grep.
What broke first was the edge typing. Naive links made every connection look equally meaningful, so the agent couldn't tell "attended" from "blocked-by." Once I forced edges to carry a type and a timestamp, the graph started telling the truth about precedence, and the agent stopped hallucinating causal relationships that weren't there.
The core issue was never storage. It was that no agent had one place it could trust to tell it the truth about who someone is and what happened last.
