SkillFoxx

Jev-Style plugin

Jev-Style runs small decision models on your machine: a probability per option, six skills, MCP and a Claude Code guard hook

  • 134 ★ on the repository
  • Last commit
  • Apache-2.0
Plugin

Extended access

Why this level

  • The guard hook reads every Claude Code tool call and can block it, so a model mistake blocks needed work or lets something through
  • Downloads weights from Hugging Face and runs a local server on `127.0.0.1:8765`
All reasons and checks

lawrence3699/jev-styleAuthor:lawrence3699

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add plugins/jev-style

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • The guard hook reads every Claude Code tool call and can block it, so a model mistake blocks needed work or lets something through
  • Downloads weights from Hugging Face and runs a local server on `127.0.0.1:8765`

This is third-party code. Review the repository files before installing.

What it does

A family of small decision models, 0.8B and 2B parameters fine-tuned from Qwen3.5, and a toolkit around them. A model takes text or JSON, optional images and typed questions, reads the input once and returns a calibrated probability for every option in one forward pass on a laptop. Question types are yes or no, multiple choice up to 255 options and ordered ratings from 2 to 10 levels. For agents the repository ships six skills, an MCP server listed in the official registry and the `jev-style guard` hook. Before a `Bash`, `Write`, `Edit`, `Read`, `WebFetch` or MCP call, the hook sends the call to the local model and asks whether it is destructive, exfiltrates data or touches secrets. The server starts with `jev-style serve` and listens on `127.0.0.1:8765`.

Who it is for. Developers of agents and automations who need fast local classification and tool-call checks with no external API.

Good fit when

  • You need to route support tickets to teams and act only when the probability clears a threshold
  • An agent calls tools in Claude Code and risky calls should stop before they run
  • The code has LLM calls that only return a label or yes and no and you want to swap them for a local model
  • You want to measure model accuracy on your own labelled data with `jev-style eval`

Not a fit when

  • You need generated text or dialogue: the models return only option probabilities
  • The hook must be the last line of defence: the author calls it a second line and no sandbox
  • You have neither 1.5 GB for weights nor Python 3.10

Example request

Start jev-style serve and classify this customer message into billing, shipping and returns with probabilities

Limitations

  • The first 0.8B start downloads about 1.5 GB of weights. Python 3.10 or newer and one backend are needed: `jev-style[mlx]` for Apple silicon or `jev-style[torch]` for NVIDIA, CPU and images.
  • Images are read only by the 2B v4 PyTorch build, while MLX and GGUF builds read text.
  • The guard agrees with hand labels on 77.6 percent of 49 calls, measured on 0.8B v3, and the 2B models were not measured on that set.
  • The project is independent and not affiliated with TypeSafe or Jev.
  • The hosted API at `api.jevstyle.com` takes text only, with an 8,192-token context.

How to disable. For the Claude Code plugin run `/plugin uninstall jev-style-guard@jev-style`, and `claude mcp remove jev-style` removes MCP. Uninstall the package with `pip uninstall jev-style`; weights sit in the Hugging Face cache.

FAQ

How do I install Jev-Style in Claude Code?

Run `/plugin marketplace add lawrence3699/jev-style`, then `/plugin install jev-style@jev-style` for skills and MCP and `jev-style-guard@jev-style` for the guard hook.

How do I connect the MCP server?

Start `jev-style serve`, then `claude mcp add jev-style --scope user -- jev-style mcp`. The MCP server uses stdio and forwards to the running server, so one model copy serves every client.

How does the guard hook check calls?

Before `Bash`, `Write`, `Edit`, `Read`, `WebFetch` and MCP calls it asks the model whether the call is destructive or exfiltrates data. Example: `jev-style guard --check "git push --force origin main"`.

Is a GPU required?

No. The 0.8B model takes 0.53 GB at 4-bit, and on an M1 Max a short request takes 0.15 to 0.2 seconds after the first call.

Can I try it without installing?

Yes, Hugging Face Spaces exist for the 0.8B and the 2B. Locally the server opens a playground at `http://127.0.0.1:8765`.

Discussion

No questions or reviews yet.

Plugins

A plugin and MCP server that hand no-text decisions to the Jev model and gate shell commands while the LLM writes content

PluginExtended access71

An MCP server giving an agent three Jev tools: typed yes or no, choice and score questions with probabilities, offline by default.

MCP serverExtended access105

A plugin with a catalogue of public Jev examples: a skill and an MCP server show how others solved the same decision task.

PluginMinimal access602

The Jev classifier as an agent tool: labels items, answers yes or no and ranks candidates with probabilities within milliseconds.

PluginExtended access59

An MCP and ACP adapter for the TypeSafe Jev decision model: typed answers and confidence gates for clicks and tool calls.

MCP serverExtended access51

A plugin for DeepSeek Harness and iPolloWork: structured choice of tools, skills and agents and scoring of outputs

PluginExtended access272

README in short

The README opens with a quick start: install a backend, start the server and make the first decision with `jev-style decide`. A model table, comparison with neighbours, six skills, the guard hook, MCP, evaluation on your data and cascades follow.

README badge

Report a problem