jkudish/jev-mcp: first run, 2026-10-11

Installed from its README in a throwaway sandbox by Claude Code (Claude Opus 5.5), with a Jev API key in the environment and nothing else: no GPU, no desktop, no other paid keys, no accounts. The README's first example was run once. Repo commit 86eae7861c60; 44 seconds. All results · Task

Quick start: ran. Jev calls made: 1.

Install	npx -y @jkudish/jev-mcp (then, per the README's Claude Code section: claude mcp add jev -- npx -y @jkudish/jev-mcp; `claude mcp list` reports jev ✔ Connected)
Quick start, as run	jev_verify, the README's first tool example: claims ["Wearing a helmet is optional for adult riders.", "The ordinance mentions reflective gear."] against the City Bicycle Safety Ordinance s.4 evidence text, called as tools/call over stdio to `npx -y @jkudish/jev-mcp` with TYPESAFE_API_KEY exported
What came back	model jev-1.13.0, provider typesafe; summary {verified:1, contradicted:1, unsupported:0, needs_review:0}; claim0 'helmet optional' -> contradicted (contradicts=1, confidence 1, action auto); claim1 'reflective gear' -> verified (supports=1, confidence 1, action auto); usage 891 input / 128 output tokens. Matches the README's 'live result, abridged'.
What stopped it	Nothing.

Steps the README does not mention

The README has no standalone quick-start command: `npx -y @jkudish/jev-mcp` only starts a stdio server and waits for input, so I took the first tool example (jev_verify) as the quick start.
To actually call a tool without an LLM-driven MCP client session, I wrote a ~20-line Node stdio JSON-RPC client (initialize, notifications/initialized, tools/call). The README does not give a CLI or script for calling a tool directly.
The env var is already named TYPESAFE_API_KEY, so I didn't need to rename it; I only had to `export` it so the spawned server inherits it.

The session's closing message

The first run worked: the project installed as the README describes, and one real Jev call returned the same result the README shows.

- **What it is:** an MCP server (`@jkudish/jev-mcp`). `npx -y @jkudish/jev-mcp` started it, and it printed `ready — model jev-latest`. I also ran the README's Claude Code command, `claude mcp add jev -- npx -y @jkudish/jev-mcp`, and `claude mcp list` then showed `jev: ✔ Connected`.
- **Quick start:** the README has no standalone demo command, so I used its first tool example: `jev_verify` checking two claims against the bicycle ordinance text. That was one Jev call. The model `jev-1.13.0` marked "helmet optional" as contradicted and "mentions reflective gear" as verified, both at confidence 1, using 891 input and 128 output tokens. That matches the README's sample output.
- **Steps the README doesn't mention:**
  - The README never says how to call a tool directly, outside an AI agent session. I wrote a ~20-line Node script that talks to the server over stdio to make the call, and didn't change any of the project's code.
  - I had to `export TYPESAFE_API_KEY` so the server could see it. The variable already has the name the project expects, so no renaming was needed.
- **Nothing blocked the run**, and the API key was never printed or written to a file.

DELIVERABLE: ~/work/output/report.json