Skip to content

local-decider

Pre-action gate and browser agent where rules decide first, then Jev or a local Jev-compatible model, emitting allow, confirm or block.

Serviceresearch

Overview

local-decider is a browser agent whose decision model, the part that picks the next action, can run entirely on a local GPU behind the same HTTP contract as TypeSafe's hosted Jev, switched by one environment variable. Its hard rule is that anything a rule can decide is decided by a rule, not by weights. A pre-action gate for a tool-using agent runs a deterministic rule chain first and only asks the model where no rule decides, emitting allow, confirm or block.

The author judges the gate by leaks rather than accuracy: a false allow is an incident, a false confirm is a question. On 112 labelled safety cases, per the listing, rules alone resolved 61 and were correct on 58, with zero false allows on irreversible tools. As a judge on 54 labelled points, hosted Jev scored 90% and a local model 85%, which is why swapping matters. The jev-ultrafast browser loop was ported to a minimal Chrome DevTools client, and task completion is verified by reading the final state from the DOM rather than trusting the agent's own report. Every number and raw log is in the repository; a Romanian README is available.

What the decision model decides

TypeQuestionAnswers
ChoiceShould this agent action be allowed, confirmed with the user, or blocked?allow · confirm · block
ChoiceWhich browser action should be taken next?—

How it works

  1. Agent proposes an action
  2. Deterministic rules decide if they can
  3. Otherwise the model (hosted Jev or local) returns allow, confirm or block
  4. Browser loop executes via CDP
  5. Final state is verified in the DOM

Strengths

  • Rules handle what they can, reducing model risk
  • Zero leaks on irreversible tools in its test set
  • Hosted or local decider by configuration
  • Evidence files for every number

Caveats

  • Local model less accurate than Jev in tests
  • Small labelled sets
  • Single GPU setup

Getting started

export TYPESAFE_BASE_URL=http://127.0.0.1:11435 TYPESAFE_API_KEY=local TYPESAFE_MODEL=winnow:e4b, then ./jev-browser doctor.

At a glance

Type
Service
Status
research
Language
Python
Runtime
browser agent with CDP and local System One server
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

  • Agent safety: a pre-action gate for a tool-using browser agent that runs a deterministic rule chain first and only asks Jev where no rule decides — anything a rule can settle is never sent to weights — emitting allow / confirm / block and swapping between hosted `jev-latest` and a local Jev-compatible port by one environment variable; on 112 labelled cases the rules alone resolve 61 and are correct on 58 (95.1%) with 0 false allows on irreversible tools, and the end state is read from the DOM rather than from the agent's own report.

    yibie/awesome-jev

Frequently asked questions

How is the gate judged?

By leaks (false allows), not overall accuracy.

Can I run it without TypeSafe?

Yes, point TYPESAFE_BASE_URL at a local Jev-compatible server.

How is task success checked?

By reading the final state from the DOM, not the agent's report.

Related projects

  • Claude Code kit wiring Jev into a tool-call guard, sub-agent sizing, file search, a browser agent, review, belay and compaction.

    ProjectbetaPython
  • Experimental browser-only MCP server where Jev picks a code-owned action ID and a human approves before execution.

    MCP serverprototypeTypeScript
  • Small one-pass classifier that labels shell commands as read-only, reversible or destructive before an agent runs them.

    ModelprototypePython
  • Local browser agent that picks each step's operation and element with a Jev-shaped model such as a fine-tuned Laya, no cloud needed.

    MCP serverbetaPython
  • Claude Desktop MCP extension that drives Brave or Chrome with Jev choosing on-page actions and a guard freezing irreversible clicks.

    MCP serverbetaTypeScript
  • Chrome MCP server with 50 tools that drives your everyday browser without losing logins, where Jev or a local decider picks each step.

    MCP serverbeta