ruLake serves verifiable retrieval over a hosted MCP endpoint

What ruLake is

Retrieval systems usually hand back a passage and ask you to trust that it is current and unaltered. ruLake fingerprints every result, so stale or tampered data is rejected rather than served, and each answer carries a decision trace.

It sits in front of a vector store or lakehouse rather than replacing it, which means you can put it in the path of what you already run.

What changed

The node-v2.2.0 release packaged the Node.js path. Since then the repository has mostly seen continuous integration refreshing live status badges for the hosted MCP endpoints, plus a May review of the Ruflo and ruLake integration surface.

Stated plainly: the most recent commits in this repository are automated badge refreshes, not feature work. This story is here for the release, not for the badges.

Get started

Pick the install path that matches where your agent runs:

cargo add rulake        # Rust
pip install rulake      # Python
npm install rulake      # Node.js / TypeScript
npm install rulake-wasm # Browsers, Workers, Deno, Bun

For Claude Code, Cursor or Cline, the plugin marketplace is the documented route:

/plugin marketplace add ruvnet/RuLake
/plugin install rulake-stack@rulake-marketplace
/reload-plugins

Expected result: the rulake-stack commands resolve, and a query returns a result with a receipt attached. The install talks to the hosted demo servers with no API keys to configure.

The reload step is required; Claude Code's own install message asks for it.

Use it today

Practical case: an agent answering questions about your architecture decision records. Input is the ADR set. Workflow is query through ruLake, read the answer and its receipt, verify the bundle. Output is an answer you can trace back to the exact document version.

Acceptance test: run a query, then verify the returned table bundle. The verification should confirm the fingerprint.

Push it further

Experimental commentary. The hosted endpoints extend to genomic and quantum tool sets alongside retrieval, which hints at a general shape: verifiable execution over domain specific data.

Limitation: the demo servers are demos, and a receipt proves integrity, not correctness of the answer. Falsifiable test: modify a source document, re-query, and confirm the receipt changes.