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

SystemOneHarness

Agent harness that compiles an environment's finite actions into typed questions for Jev and gates each step by confidence.

CLIbetaApache-2.0
View on GitHub ↗

Overview

System One Harness turns a System One decision model into an agent loop. It observes an environment, compiles the finite set of available actions and parameter values into typed choice, noul and score questions, asks the model once per step, checks the weakest required probability against a risk threshold for that action, executes it and records the full trace. No action text is generated, and every transition carries a probability.

Jev, through OpenRouter or TypeSafe directly, is the first supported model. Environments can be Python processes, MCP servers or web pages, including a live browser game shown in the README. Read, write and destructive actions can have different thresholds, and every run ends as completed, incomplete, failed or cancelled with a structured reason. A built-in order-fulfilment example finishes in five steps for a fraction of a cent. The project includes a test suite and reports passing a UHP core conformance check, and it is aimed at developers who want bounded, auditable agents rather than free-form tool calling.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich declared action should be taken next, with which parameter values?—
Yes / noDo the action's preconditions hold in the current state?(inferred)—

How it works

  1. Observe the environment state and its available actions.
  2. Compile actions into typed choice, noul and score questions.
  3. Ask the model once per step.
  4. Gate: the weakest required probability must clear the action's risk threshold.
  5. Execute, record the trace and loop until a terminal outcome.

Strengths

  • The model can only pick declared actions.
  • Risk-tiered confidence gates for read, write and destructive actions.
  • Complete step-by-step traces with latency and cost.

Caveats

  • Needs an OpenRouter or TypeSafe key for the live model.
  • Bounded action spaces only; open-ended tasks need an environment adapter.

Getting started

pip install -e . then export OPENROUTER_API_KEY (or TYPESAFE_API_KEY) and run s1 run --env order:ship_fastest_gift.

At a glance

Type
CLI
Status
beta
Language
Python
Runtime
CLI + MCP/browser adapters
Has code
yes
Has tests
yes
License
Apache-2.0
Stars
193

Use cases and integrations

How curated lists describe it

  • Model evaluation: execution harness and dual-loop test framework that compiles goals, browser environments, and MCP servers into bounded System One reflexes, evaluating Jev against deterministic baselines.

    yibie/awesome-jev

  • The System One harness for System One models. Run Jev and other System One models locally or directly on HarnessRouter.ai.

    wh000wh000/awesome-jev-live

  • Turns a decision model into an agent loop: it compiles the environment's finite action space into typed questions, gates each step by confidence, and records the trace.

    MrJev/awesome-jev

Frequently asked questions

What does the model decide each step?

Which declared action to take and with which parameter values, chosen from typed questions compiled from the environment.

What happens if the model is unsure?

The step is gated: if the weakest required probability is below the action's threshold, the run stops with a structured reason instead of acting.

Which environments are supported?

Python processes, MCP servers and web pages through optional extras.

Related projects

  • jev-judgment

    Agent skill that sends closed coding-agent judgments to Jev before asking the user, before risky commands and after failures.

    ModelbetaPython
  • AskJev

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

    MCP serverbetaTypeScript
  • BrowserPaw

    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
  • jev-adb

    MCP server that lets agents observe and control Android phones over ADB, with Jev picking controls and judging goal completion.

    MCP serverbetaTypeScript
  • jev-mobile

    Durable Android sub-agent: a SQLite-backed worker observes the UI and Jev picks among valid actions, with verification and escalation.

    MCP serverbetaPython
  • macos-computer-use-kit

    Accessibility-first macOS computer-use MCP server and CLI, with optional Jev guards on targets and inputs before irreversible actions.

    PluginbetaPython