Spots Fast, typed decisions for agent harnesses. Local gates and receipts for every choice. Route tasks, tools, skills, tests, and reviews through TypeSafe Jev, with optional local Laya-MLX on Apple Silicon. Not every decision inside a coding agent needs another chat-model turn.
Qualixar Jev Decision Layer is an open-source MCP decision layer for AI agent harnesses. It routes bounded choices—task routing, tool selection, skill choice, file ranking, test selection, and review priority—to hosted TypeSafe Jev or optional local Laya-MLX. A local policy gate checks the typed answer and records a receipt. The agent's host keeps permission to execute and verify the work.
Why Jev for the decision step? TypeSafe's published System One workflow comparisons report Jev at 193.6× faster and 444.6× cheaper than the LLM reference on those workflows. Its model pricing lists $0.042 per million input tokens, with output tokens free. Qualixar connects that fast decision primitive to reusable agent recipes, local gates, workspace controls, and evidence you can inspect.
One shared runtime has adapters for Codex, Claude Code, VS Code, Hermes, and Antigravity. Those adapters do not have equal live-verification evidence; the host table states what has and has not been run.
Why use a layer instead of calling Jev directly? The model call is one step in an agent workflow. This repository adds reusable decision contracts, workspace scope and request controls, a local answer gate, receipts, optional local routing, and host-specific adapters behind one shared runtime.
The model recommends; the host retains execution authority. A Jev or Laya answer does not grant shell, file, browser, deployment, or publishing permission. Host support is not equally verified; see the evidence and boundary table.
Current release: 1.0.8 — 20 MCP tools, 38 recipes for bounded decisions, 114 synthetic offline fixtures, and five host adapters. The live recipe gate caps every would-be passing result to verify until a recipe has been evaluated against labeled provider answers. Version 1.0.8 is supported and verified on macOS. Linux remains experimental and unverified; Windows hosted operation is disabled in this release because its native private-state contract did not pass CI. Local Laya-MLX is limited to supported Apple-Silicon Macs. See the host evidence and changelog.
Built by Varun Pratap Bhardwaj under Qualixar. This is an independent open-source integration, not an official TypeSafe, OpenAI, Anthropic, Google, Microsoft, or Laya product. Install the plugin, complete the reviewed workspace setup, then ask your agent:
Route this synthetic duplicate-charge request between billing and engineering. Show the selected candidate, provider, model, route status, and receipt ID.
The prompt supplies a closed choice. A live result identifies the provider/model, route status, and receipt; the selected candidate may vary. This route example does not return the recipe-specific host_action. The animation in What can an agent actually do with it? illustrates the answer flow. Why an AI agent decision layer instead of a direct Jev call?
Fast, typed decisions for agent harnesses.