Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

JevRouter

Router that puts models, subagents, skills, MCP tools and CLIs in one candidate set and asks Jev one Choice about which handles the step.

MCP serverbetaMIT
View on GitHub ↗Live demo ↗

Overview

JevRouter targets the question agents keep spending reasoning tokens on: which capability should handle this next? It merges models, subagents, skills, MCP tools, CLIs and plugins into one candidate set, asks Jev a single typed Choice, and then enforces availability, permissions, risk and confirmation around the answer. Jev owns the probabilities; the router owns policy. Filtered candidates are never re-normalised, and the original probabilities and provider response are kept.

It is decision-only by default: nothing executes implicitly, and medium, high or critical capabilities need confirmation. A route call answers one question, while plan answers which capability handles steps 1 to N with serial, batch or decomposed strategies. Below the configured confidence it returns no_decision. Decisions and plans are written as append-only receipts with provenance hashes, and strict projects can require verified candidate descriptions. On first-five tool-call prediction over 10 Toolathlon tasks, the README reports 38 to 44% position-wise hits for Jev against 24% for DeepSeek V4.1 Flash, at lower cost.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich capability (model, subagent, skill, MCP tool, CLI or plugin) should handle this step?—

How it works

  1. Build the candidate set from configured and discovered capabilities
  2. Jev answers one Choice over the candidates
  3. Router filters by availability, permissions and risk; requires confirmation for risky ones
  4. Below the confidence threshold, return no_decision with original probabilities
  5. Write an append-only receipt with provenance hashes

Strengths

  • One routing contract across many capability types
  • Nothing executes implicitly
  • Auditable receipts
  • Benchmark against an LLM baseline published

Caveats

  • Benchmark covers only 10 tasks
  • Measures routing, not end-to-end task success
  • Unofficial; needs a Jev key

At a glance

Type
MCP server
Status
beta
Language
TypeScript
Runtime
Node library, CLI and MCP server
Has code
yes
Has tests
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Coding agents: asks Jev a typed `Choice` over models, subagents, skills, MCP tools, CLIs and plugins, applies availability, permission and risk policies, and returns `no_decision` below the configured confidence threshold while preserving the original probabilities.

    yibie/awesome-jev

  • Unofficial router that puts models, subagents, skills, MCP tools, and CLIs in one candidate set: Jev answers one Choice, and code enforces availability, permissions, risk, and confirmation. On 10 Toolathlon tasks, position-wise hits were 38–44% for Jev against 24% for DeepSeek V4.1 Flash

    AnotiaWang/awesome-decision-models

  • Routes among models, subagents, skills, MCP tools and CLIs using a shared candidate set.

    logicrw/awesome-jev-projects

  • A lightweight Jev-powered router for models, tools, and subagents.

    wh000wh000/awesome-jev-live

  • Routes agent requests across models, subagents, skills, MCP tools, CLIs, and plugins with Jev Choice decisions; the host filters by availability, permissions, risk, and confirmation before anything executes.

    cobanov/awesome-jev

  • A lightweight Jev-powered router for models, tools, and subagents.

    hellogumbo/awesome-jev

  • Routes each agent step to a model, subagent, Skill, MCP tool, or CLI with one Jev Choice, requiring confirmation for risky capabilities and keeping decision receipts.

    MrJev/awesome-jev

Frequently asked questions

What does Jev decide?

Which capability in the candidate set should handle the current step, with a probability per candidate.

Can it execute tools?

It is decision-only by default; execution is left to the host, with confirmation for risky capabilities.

How well does it route?

On 10 Toolathlon tasks, 38 to 44% position-wise hits versus 24% for DeepSeek V4.1 Flash, per the README.

Related projects

  • Helm

    Orchestrator that probes which coding agents are installed and logged in, then has Jev pick the best one for a task.

    Agent skillbetaPython
  • jev-agent-hooks

    Hooks for Claude Code, Codex and pi that use Jev to suggest a skill each turn, plus a rule gate for subagent model routing.

    HookbetaJavaScript
  • jev-pilot

    Claude Code plugin that asks Jev before each turn for reasoning effort, subagent model, strategy and skill, with a decision log.

    Agent skillbetaTypeScript
  • jev-for-all

    Experimental plugin wiring Jev into coding-agent harnesses to pick skills, tools, models and browser moves, starting with OpenCode.

    Agent skillprototypeTypeScript
  • todo-jev

    Experimental task router that uses Jev, skill profiles and environment checks to send requests to a rule, a skill or a large model.

    Agent skillprototypePython
  • jev-harness-router

    Per-turn router: one ~350 ms Jev call picks model tier, effort, tools and skill, with a hard deadline and regex fallback.

    Agent skillbetaTypeScript