Why add Jev when Kinde permissions already gate the API?

Permissions say what the user may do; Jev asks whether the user actually asked for this call. It scores match, destructiveness, planted-text following, exfiltration, and risk, catching prompt-injected calls permissions allow.

Does every read pay model latency and cost on the hot path?

Reads skip Jev when permitted, so they add no cost. Other calls go to Jev in about 200 ms, with low-confidence calls falling to an LLM judge; ledger-first writes mean a failed write never runs.

What happens to a destructive call that needs a human decision?

Held calls wait 10 minutes for approval on a page requiring fresh sign-in, then run once. The demo ledger, console, attack playground, benchmark, and connect pages show the flow end to end.

Are the benchmark wins over single-verdict and LLM judges real?

The kit reports 0.0% false-allow at 83.7% accuracy against 12.4% and 7.9% false-allow for single-verdict and LLM baselines on 300 labeled calls run 3 times each. Re-run on your deployment to confirm.

What do I need in order to run Jev Gatehouse?

Node.js 20.17 or later, a Kinde business, a Convex account, and an OpenRouter API key, which is used for Jev, the LLM judge and the in-app agent.

How does the guard handle a destructive call?

Jev returns calibrated signals and code decides. A call that needs approval is held for 10 minutes; the user approves it on a page that requires a fresh sign-in, and the held call then runs once.

What do the benchmark numbers show?

Over the same 300 labeled calls, Jev signals with policy code had a 0.0% false-allow rate at 205 ms and $0.029 per 1,000 calls, against 12.4% for a Jev single verdict and 7.9% for an LLM judge at 2,073 ms and $0.931.

CommunityJev Practices & Patterns

Jev Agent Authorization

Starter kit putting Kinde identity plus Jev judgment in front of every agent tool call, with allow, hold-for-approval, or deny decided in code and written to a ledger first.

See highlights

Open github.com

Stars: 5
Language: TypeScript
License: MIT
Published by: Community
Author: kinde-starter-kits
Last verified: Oct 2026

kinde-starter-kits/jev-agent-authorization

Organization repository on GitHub

View on GitHub

Jev agent authorization for MCP tool calls: Kinde identity and permissions plus Jev's typed, calibrated decisions, checked server-side before every call runs

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

A starter kit that checks every AI agent tool call before it runs. Kinde checks who the agent acts for and what that user can do. Jev, the System One model from TypeSafe AI, judges each call in about 200 ms. Policy code in the app decides: allow the call, ask the user to approve it, or stop it. An agent calls a small workspace API through a Kinde MCP connection. The Kinde MCP server passes the user's access token to the API. The guard in front of the API runs these steps for each call:

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Verifies Kinde tokens, permissions, and flags before any Jev call or workspace write.
Sends call plus request plus read documents to Jev for calibrated multi-signal output.
Decides allow, step-up approval, or deny in policy code, not in model prose.
Streams every decision with checks, signals, and cost to a live ledger and benchmark page.

Watch out

Benchmark labels are author-written and should be remeasured on your own deployment with npm run bench before trusting the ranking. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-codes

CLI and agent commands that audit a git diff hunk-by-hunk against editable YAML standards packs, with Jev answering typed questions and the agent fixing flagged lines.