OpenClaw Jev gate

It gives an OpenClaw bot two Jev-powered decisions before the language model runs: a reply gate that decides whether a group message deserves an answer, and a model router that decides which of up to three configured models should handle it. Each feature has its own enabled switch.

requireMention true answers only when named, and false answers every message. The gate is the middle mode: the bot sees every message and answers the ones meant for it — mentions always, follow-ups right after the bot replied, while chatter like lunch plans or thanks stays silent and never reaches the model.

Jev estimates how much work a request takes and whether a wrong answer has consequences, then picks light, standard or heavy from the models you configure. Unknown requests go to standard, never to light; a tier left empty keeps the agent's own model, and with no tiers set the router only logs what it would have done.

The reply gate sends the text of each group message plus the last few messages of that conversation (8 by default, each cut to 400 characters) and the assistant's name and description; direct messages are not sent unless evaluateDirectMessages is on. The router sends the last 4000 prompt characters. Nothing is sent for cron, heartbeat, memory or overflow runs.

Only agents that OpenClaw runs itself against a provider API, such as anthropic/... or openai/... with a key. It does nothing for agents on a CLI runtime like claude-cli, where the run is handed to an external CLI; the README labels this prototype status explicitly.

The plugin fails open: if Jev is unreachable, slow or returns an error, the bot behaves as if the plugin were not installed — the gate lets the message through and the router keeps the agent's model. The API key is never logged, and OpenClaw requires an explicit permission before the router reads prompts.

The README calls it a prototype. The reply gate runs on a live gateway. The model router loads on a live gateway and its questions were checked against the live Jev API with nine prompts, but it has not yet routed a real run because the test gateway's agent runs on Claude CLI, which the router does not support.