jev-x-kit
MCP server, CLI and Claude Code skill with Choice, Score and Noul primitives and a confidence gatekeeper, hosted or offline.
Overview
jev-x-kit packages a System One decision layer for coding agents as a Claude Code plugin and skill, an MCP server with 21 tools, and a standalone CLI. It targets problems the author lists for agent loops: slow LLM yes/no routing, token cost on micro-decisions, broken JSON, hallucinated summaries, single-pass blind spots and vendor lock-in. Primitives are Choice (up to 255 options), Score and Noul, schema-validated and never free text, with speculative fan-out of independent questions in one pass.
A confidence gatekeeper executes directly above 0.85, splits into sub-decisions between 0.60 and 0.85, and escalates to a larger model or a human below 0.60. The kit also offers planning, red-teaming, research and self-improvement features. Backends are selected automatically or explicitly: TypeSafe Jev's native systemone API, OpenJev on vLLM, local Laya, Ollama for the larger planner model, or a heuristic offline simulator that keeps every primitive schema-valid with no key or GPU. Note that the free offline mode is a deterministic simulator, not a decision model. It has CI and 27 unit tests.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Agent micro-decisions posed as Choice, Score or Noul (routing, gating, review) | — |
How it works
- Agent calls a primitive via MCP, CLI or skill
- Backend answers (Jev, OpenJev, Laya or offline simulator)
- Gatekeeper acts above 0.85, splits 0.60-0.85, escalates below 0.60
- Optional planning and red-team passes use a larger model
Strengths
- Several backends including fully local options
- Clear confidence bands for escalation
- CI with unit tests
Caveats
- Offline mode is a heuristic simulator, not a real model
- Broad feature set with ambitious claims not benchmarked in the README
Getting started
git clone https://github.com/Kadihx/jev-x-kit.git && cd jev-x-kit && npm install && npm run build, then load it as a Claude Code plugin.At a glance
- Type
- MCP server
- Status
- beta
- Language
- TypeScript
- Runtime
- MCP server / CLI / Claude Code plugin
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Offline $0 decision layer for coding agents: Choice/Score/Noul primitives…
wh000wh000/awesome-jev-liveOffline $0 decision layer for coding agents: Choice/Score/Noul primitives, BELKI confidence gatekeeper, ultra-planning, red-teaming, research and RLVR self-improvement -- as an MCP server + CLI + Claude Code skill.
hellogumbo/awesome-jev
Frequently asked questions
Does it need an API key?
No for the offline simulator; real decisions need TypeSafe Jev or a local OpenJev or Laya backend.
How are uncertain answers handled?
Above 0.85 runs directly, 0.60-0.85 splits into sub-decisions, below 0.60 escalates.
Which agents does it target?
Claude Code, Cursor, Codex, OpenCode, Continue and any MCP client.
Related projects
Rust engine that answers noul, choice and score from any LLM in one prefill via logprobs, wire-compatible with /v1/systemone and MCP.
ModelbetaRustJev decision layer for agents: seven MCP tools, an embeddable library and escalate-only Claude Code hooks at harness boundaries.
MCP serverbetaTypeScriptNative Bend executable that owns agent-launched processes and logs, and asks Jev a typed question only when told to.
ServicebetaBendLocal MCP server putting Jev in the coding loop: route next steps, pick prepared tool calls, review patches and verify claims.
MCP serverbetaTypeScriptMCP gateway that trims noisy tool output and offers task classification and review advice, splitting work between local Laya and Jev.
MCP serverprototypeTypeScriptOpen-source dual-brain runtime that answers typed decisions with a fast local engine or the Jev API before falling back to LLMs.
ServicebetaPython, TypeScript