Any markdown vault in, any MCP-capable agent out. SQLite in the middle. No cloud, no lock-in, no proprietary format.
The indexer parses YAML frontmatter, [[wiki links]], and heading-path
chunks — the common denominator of every markdown PKM tool. No plugin required.
The reference target. Wiki-links, frontmatter, and Dataview-style inline fields all index cleanly.
PKMOutliner-style markdown with block references. Works as a plain folder of .md files.
VS Code-native knowledge base. Same markdown + wiki-link contract the indexer expects.
Anything else that writes plain .md with
frontmatter works too — the engine has no vault-app dependency.
The MCP protocol means any compliant agent queries the same
vault-index.db. bootstrap.sh auto-detects and installs skills into each.
Primary target. Register vdb serve in .mcp.json; skills install to ~/.claude/skills/.
Open-source coding agent. Shares the skills hub via symlink.
Detected by bootstrap.sh automatically. Shares the same skills hub and MCP server.
Another MCP-capable consumer — queries the same vault-index.db through vdb serve.
MCP-capable editor — points at the same server, same tools.
Detected by bootstrap alongside the others; per-harness model routing via vdb agents.
Agents run on a model_tier set by who catches the mistakes: fixed-shape tasks
(notes, commits, status flips) run tier 1; work you review runs tier 2; planning, review,
and stacks you're still learning run tier 3 — there the model is its own safety net.
Declare expertise with vdb stacks set; vdb models detects what's
available per harness; vdb agents --detect writes the agent files (role suite,
per-stack developers, and delegate-marked skills).
Raw notes are immutable source; the DB index is the compiled binary the agent works against. The agent is the programmer, the vault the codebase, the index the build artifact.
PracticePlain files you own, linked by [[wiki links]], readable forever. This project keeps
that contract and adds a query layer on top — never replacing the notes.
Agents don't pull you into their tools — you work from the vault and reach out: planning & delivery, code repositories, infrastructure, and remote services all sync through explicit, user-controlled connections. You own the data. It adapts to your workflow, not the other way around.