@vibeleading/mcp-hybrid-data-engine
v0.1.1
Published
MCP server implementing the Vibe Leading Hybrid Data Engine (Stone + Light + Flow).
Maintainers
Readme
@vibeleading/mcp-hybrid-data-engine
The Memory Hub.
An MCP server implementing the Vibe Leading Hybrid Data Engine — the integration of the three data planes described in Vibe Leading The AI:
| Plane | Role | This server | Production backend |
|---------|---------------------------------------------------|------------------------|-----------------------------|
| Stone | Legacy SQL / Oracle, fact ledger | query_stone (SQLite) | Oracle / Postgres + CDC |
| Light | Vector semantic context | index_light / semantic_search | Pinecone / Milvus + embeddings |
| Flow | Analytical columnar for the HUD (milliseconds) | flow_put_rows / flow_query | DuckDB / Parquet |
Installation
npm install -g @vibeleading/mcp-hybrid-data-engine
# or run directly without installing:
npx -y @vibeleading/mcp-hybrid-data-engine
# or run straight from the GitHub source (builds automatically):
npx -y github:VibeLeading/mcp-hybrid-data-engineRequirements
- Node.js >= 22 — required. The Stone relies on the built-in
node:sqlitemodule, which does not exist in older Node versions. If the server cannot start (missing build or old Node), MCP clients report a generic "connection closed" error. - No other runtime dependencies — everything ships with the package.
Client configuration
Add to your MCP client config (e.g. Claude Code .mcp.json or Cursor mcp.json):
{
"mcpServers": {
"hybrid-data-engine": {
"command": "npx",
"args": ["-y", "@vibeleading/mcp-hybrid-data-engine"]
}
}
}Tools
| Tool | Description | Key args |
|--------------------|--------------------------------------------------------|-------------------------------------------------|
| query_stone | Read-only SQL against a SQLite file (Stone) | sql, database? |
| stone_cdc_tick | CDC version + changed-tables heuristic | — |
| index_light | Add a doc to the semantic index (Light) | id, text, metadata? |
| semantic_search | Ranked cosine-similarity matches | query, top_k?, threshold? |
| flow_put_rows | Store columnar rows for analytics (Flow) | table, columns, rows |
| flow_query | sum/avg/min/max/count with optional grouping | table, operation, column, groupBy? |
Environment variables
| Variable | Default | Purpose |
|-------------------|-------------------|------------------------------------------|
| MCP_STONE_PATH | data/stone.db | SQLite file used by The Stone |
| MCP_LIGHT_PATH | data/light.json | Persistence file for The Light index |
| MCP_FLOW_PATH | data/flow.json | Persistence file for The Flow store |
Notes
- Stone: read-only by design. Statement prefixes are validated; anything
outside SELECT/PRAGMA/WITH/EXPLAIN is rejected. The
stone_cdc_tickreports a coarse WAL-based heuristic — a production deployment should wire a real CDC adapter (Debezium + Oracle) behind the same surface. - Light: uses a deterministic bag-of-character-ngrams hashed-vector
embedding with cosine similarity. Swap in a real embedding provider
(OpenAI / voyage) and a real vector DB (Pinecone / Milvus) behind the
Embedderinterface insrc/vector-store.ts. - Flow: an in-memory columnar store persisted to JSON. See
src/analytics-engine.tsfor the interface that a DuckDB / Parquet backend would implement.
Publishing
Published to the npm registry via OIDC trusted publishing — no npm tokens stored
anywhere. Pushing a version tag triggers the .github/workflows/publish.yml workflow:
npm version patch -m "release: v%s"
git push origin main --follow-tags- One-time bootstrap. The very first published version cannot use trusted
publishing (the trust relationship attaches to an existing package). Publish
0.1.1once manually (npm publishafternpm login, or a short-lived publish-scoped token), then configure Trusted Publisher on npmjs.com → the package → Settings → Trusted publishing → GitHub Actions → ownerVibeLeading, repomcp-hybrid-data-engine, workflowpublish.yml, actionallow npm publish(requires 2FA once per package). - Requires Node.js >= 22 and npm >= 11.5.1 on the runner (handled by the workflow).
License & Attribution
MIT — Copyright (c) 2026 Jean Machuca (see LICENSE).
This server implements concepts from the book Vibe Leading The AI: The Corporate Race Against Machines by Jean Machuca (ISBN 9798252505008, © 2026 Jean Machuca). The book is copyrighted commercial material; this repository does not republish its text. Buy the book at https://vibeleading.org or https://a.co/d/04L5YatK. Author website: jeanmachuca.com · Support on GitHub Sponsors: github.com/sponsors/jeanmachuca. See NOTICE.
