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

reflex-gate

Windows desktop gateway wrapping a local Qwen2.5-Coder 1.5B for agent loop-exit, extraction and scan checks plus a /v1/systemone shim.

Serviceprototype
View on GitHub ↗

Overview

ReflexGate wraps a compact local model, Qwen2.5-Coder-1.5B-Instruct on llama.cpp, into a low-latency gateway at http://127.0.0.1:8090 for autonomous agent loops. It ships a Windows desktop GUI (Wails, Go) that downloads the model and a matching llama-server build, plus a performance telemetry display. The pitch is avoiding the 2-5 s and token cost of asking a cloud model whether a loop should stop.

Three endpoints take a log payload: /jev/stop for loop-exit gating (a one-line chain of thought constrained by GBNF grammar), /jev/extract for tool-argument and error-code JSON extraction, and /jev/scan for regex secret detection plus semantic crash and prompt-injection scanning. A /v1/systemone endpoint accepts Jev-shaped noul and choice questions for local triage before a cloud worker. The README is explicit that this is not TypeSafe's Jev, not a distillation, and may lose to the commercial API on hard cases; all log content stays on the machine.

What the decision model decides

TypeQuestionAnswers
Yes / noShould the agent loop stop now?—
ExtractExtract tool arguments, status and error codes as JSON—
Yes / noDoes this log contain secrets, a crash or prompt injection?—

How it works

  1. Launch reflexgate.exe and auto-fetch the model and llama-server
  2. Agent posts its latest log to /jev/stop, /jev/extract or /jev/scan
  3. Local model answers with grammar-constrained output
  4. Agent stops, continues, or escalates to a cloud model

Strengths

  • Runs offline with zero token cost
  • Sub-second local responses per the README
  • Log content never leaves the machine

Caveats

  • Not TypeSafe's Jev; hard cases may be worse
  • Windows-focused GUI
  • Self-test accuracy is on its own small suite

Getting started

Download reflexgate.exe from GitHub Releases, click Auto Fetch, then Start Server; the gateway listens on http://127.0.0.1:8090.

At a glance

Type
Service
Status
prototype
Language
Go
Runtime
Windows desktop app + local HTTP server (llama.cpp)
Has code
yes

Use cases and integrations

How curated lists describe it

  • Local Jev / System One–compatible gateway — not TypeSafe’s Jev. Fast, privacy-first, offline.

    hellogumbo/awesome-jev

Frequently asked questions

Is ReflexGate the same as Jev?

No; it is a local Qwen-based gateway with a Jev-compatible endpoint shape, not affiliated with TypeSafe.

What does it decide?

Whether an agent loop should stop, what error or argument data to extract, and whether a log contains secrets or injection.

Does it need an API key?

No; everything runs locally.

Related projects

  • jev-x-kit

    MCP server, CLI and Claude Code skill with Choice, Score and Noul primitives and a confidence gatekeeper, hosted or offline.

    MCP serverbetaTypeScript
  • AgentJev

    Open-weight 0.6B decision model on a Qwen3 backbone: state and typed questions in, a probability for every option out, nothing decoded.

    ModelresearchPython
  • `kaijia323/dsh-plugin-jev`：DeepSeek Harness 的 Jev 插件（与本库同环境）

    DeepSeek Harness plugin that registers a native jev_decide tool and declares its own policy for when the agent should ask Jev.

    PluginbetaJavaScript
  • ask-jev-skill

    Hermes skill that calls Jev as a typed tiebreaker, returning Choice, Score or Noul probabilities so the agent can act or escalate.

    Agent skillbetaPython
  • decidealot

    Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.

    MCP serverbetaPython
  • decision-gate

    npm library that rate-limits, spend-caps, caches and redacts every Jev request made in a loop.

    LibrarybetaJavaScript