Jev Tool Permissions

Before a tool call executes, a pre-flight approval gate decides auto-approve, ask-human or block. Before a turn, tool-list pruning drops tool definitions irrelevant to the current task - useful when agents carry 50 to 200 MCP tools costing thousands of tokens every turn.

The gate runs on every single tool call, on the critical path. The README argues a frontier chat model would be slower and many times more expensive than the call it gates, while Jev returns typed probabilistic answers - probability, score, or choice with probabilities - so thresholds need no output parsing.

Both features are fail-closed. If the Jev call errors or times out, the approval gate defaults to ask-human, never auto-approve, and pruning keeps everything rather than silently dropping potentially relevant tool definitions. The README calls out this behaviour explicitly.

npm install jev-tool-permissions ai, where ai ^7.0.105 is a peer dependency and no other runtime dependencies are added. Authentication comes from AI_GATEWAY_API_KEY, or a roughly 12-hour VERCEL_OIDC_TOKEN when running inside a Vercel project.

Both the gate and the pruner sit on allow and deny rules that run first and never call the model, modelled on how fast-jev-compaction uses pinning. Rules can be globs, regular expressions, or predicates over the tool call's arguments, and deny rules are checked first.

The default model id is the string 'typesafe-ai/jev', which resolves through the Vercel AI Gateway. If neither AI_GATEWAY_API_KEY nor VERCEL_OIDC_TOKEN is set, createPermissionGate and pruneTools fail fast with a JevAuthError instead of surfacing a confusing 401 from inside evaluate().

No. Like the rest of the AI SDK's AI Gateway integration, it reads credentials from the environment. If you pass your own EvaluationModel instance instead of the default string id, the auth check is skipped because you are assumed to have wired up your own credentials.