LEEVAR reliability battery

leevar-mcp

The LEEVAR reliability battery as an MCP server. Ask Claude, Cursor or any MCP client to grade an AI agent's conversations on 18 reliability tests, without leaving the editor.

It does not grade anything itself and it never fills a gap with a guess. A dimension your samples cannot evidence comes back NOT TESTED, and below 13 of 18 evidenced tests no grade is issued at all — the tool says so instead of printing a letter.

Tools

Install

Node 18 or later. The package is installed straight from the v0.1.0 release; it is not on the npm registry yet.

Claude Desktop (claude_desktop_config.json) or Cursor (.cursor/mcp.json):

{  "mcpServers": {    "leevar": {      "command": "npx",      "args": ["-y", "https://github.com/forevercrab321-svg/leevar-battery/releases/download/v0.1.0/leevar-mcp-0.1.0.tgz"]    }  }}

Claude Code:

claude mcp add leevar -- npx -y https://github.com/forevercrab321-svg/leevar-battery/releases/download/v0.1.0/leevar-mcp-0.1.0.tgz

From source instead: clone this repository, run npm install in packages/mcp, and point your client at node <clone>/packages/mcp/index.mjs.

Then ask, for example: "List the LEEVAR battery, pick three of my support bot's conversations that cover it, and scan them."

Your data

Samples go to LEEVAR's hosted service, are graded there by third-party language models (named on the privacy page), and are deleted 30 days after the scan completes. Each scan uses one of the email's 5 free monthly scans. Strip API keys, credentials and personal data before sending. The tool descriptions tell the assistant to ask you before it sends anything.

To grade entirely on your own machine with a model you choose, use the scanner in the root of this repository instead (deno task scan) — nothing leaves your machine except the calls to your own model provider.

Reading a result

• Grade issued — Grade: B (composite 81.2) with the coverage it was computed on. Gate on coverage as well as the letter.
• No grade — the samples did not evidence enough of the battery. Send longer, real conversations that exercise the missing behaviours: a long thread for context, a repeated task for consistency, a tool failure for recovery. A reference average may be shown; it is not a verdict.
• Failed — the reason, whose fault it was, and whether a retry makes sense. If the fault is LEEVAR's, the free scan is credited back.

Develop

npm installnpm test          # all requests go to local fakes; no free scans are spentnpm run gen       # regenerate battery.json from packages/scanner-core (Deno)

LEEVAR_API_URL overrides the endpoint (the tests use it to point at a local stub).

MIT.