evaluate( )

Pattern

Jev Tool Permissions

Prune an agent's tool list before the model sees it. Allow and deny rules run first; Jev asks one relevance boolean per remaining tool in a single call.

IntermediateSep 2026View the full pattern on desktop
Jev Tool Permissions preview

Problems solved

  • Stop paying for irrelevant tool definitions on every turn
  • Keep dangerous tools out of the model's view with deterministic rules
  • Prune a large MCP tool set without dropping tools the task needs

Use cases

  • Trimming a 50–200 tool MCP set per task
  • Allow and deny rules for agent tools
  • Teaching Jev batched booleans with prepareStep

AI SDK APIs

experimental_evaluateexperimental_filterActiveToolsprepareStep

Tags

jevevaluationtyped-decisionsbooleantoolsmcpprepareStepactiveToolsai-sdk

Requirements

Node.js
24+
Memory
256MB
Other
  • Vercel AI Gateway API key

Environment variables

External services

vercel-ai-gatewayupstash

Ready to use?

View the full interactive pattern with live preview and copy the source code.

View the pattern on desktop

Related patterns

Jev Email TagsTag a closed inbox with one Jev boolean per tag. A stamp lands only when its probability clears a floor.Sep 2026
Jev Table QueryFilter or sort a closed table with one Jev question per row. A column name never leaves the app.Sep 2026
Jev Semantic ClusteringDiscover themes in comments with no labels up front. Embeddings pick likely pairs, Jev answers one boolean per pair, and code does the grouping.Oct 2026
Jev Entity ResolutionResolve messy card statement descriptors to the merchants behind them. Code blocks and vetoes, Jev answers one boolean per pair, and unsure matches wait for review.Oct 2026