Jev Harness

Agent arena

Settings

Jev API key

Set a manual TypeSafe API key override for comparisons in this browser.

Active key sourceServer default

Uses the server key when configured. Save your own key below to override it.

Stored unencrypted on this device for this origin. The saved key is never shown again or included in results. Only explicit live requests use it.

Usage & budget

This tab's latest 200 logical Jev calls · survives refresh. Account-wide quotas and balances are unavailable.

Logical calls0
Reported input0
Reported output0

Physical requests: 0 · 0 observed (lower bound when unknown) · 0 additional requests observed.

Run a comparison to see Jev usage. Agent CLI usage is shown in each result.

Reported-input price estimate$0.00000000

At $42 per billion input tokens (checked Sep 22, 2026). Excludes output, unknown usage and account-specific pricing. Not an invoice.

Request history
Provider-reported tokens. Missing usage stays unknown.
TimeStatusInputOutputDurationAttempts
Request limits

Local host: 30 physical requests/minute, 2 logical calls concurrent across keys and tabs. Local safeguards, not account quotas. Billing (402) and rate-limit (429) failures remain visible. The recorded recovery policy permits either one request or up to three for probability-sum defects only. Valid answers and all other failures stop.

Jev Harness / Agent experiments

One task. Two tool menus.

Compare Codex with and without Jev routing. Keep the evidence and watch performance over time.

Stored on this device
Independent community projectSynthetic fixtures onlyEvidence, not authorizationTypeSafe UI ↗
Choose an example
Task preview

Read src/sum.ts in the synthetic workspace. Explain what sum([2, 3]) returns, citing the file.

Ready to compare.

Agent arena

How the comparison works

One logical routing call with all physical requests recorded, then two parallel Codex CLI runs using your local sign-in. Without Jev sees all fixture tools; With Jev sees the selected tools plus any reading prerequisites. Both use the same synthetic task.

Read-only settings. Proposed patches are recorded, never applied. The inspector tool is deterministic. Independent model choices and cache effects can change results; no answer-quality score is inferred.

Without Jev

The full tool catalog

Codex gets every fixture tool.

With Jev

A selected tool menu

Jev routes first. The host keeps the reading tools needed for a grounded proposal.

Run an example to compare the answers, calls, input tokens and timing. Each result is saved locally for your next visit.