system1-agents
Prebuilt browser, computer-use, robotics and game agents whose actions come from Jev, Laya or Cua-S1, with a scaffold for new agents.
Overview
system1-agents (s1a) provides ready-to-run agents whose every decision is made by a System One model instead of a chat model: Jev, Laya, Cua-S1 Nano, or OmniJev for screenshot-based browser decisions. Shipped agents cover browser use on live sites, Windows computer use, ALFWorld household tasks, a ticket router and games such as 2048, Blackjack and Millionaire. A scaffold helps build new agents, and Claude Code or Codex can run or create agents through a plugin, a browser subagent and an MCP server.
The README's benchmark table compares Jev with a chat model on each scenario; for example it reports a WebVoyager Allrecipes task taking 35.7 s with Jev versus 138.7 s with the chat model, and summarizes results as up to 6 times faster and 25 times cheaper at the same score. A decide command prints one JSON object with the choice, per-option probabilities, confidence and timing.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action should the agent take next (element and operation, move, or game action)?(inferred) | — |
| Choice | Which queue should this ticket go to? | — |
How it works
- Describe the task or pick a prebuilt agent.
- The agent turns observations into state and legal options.
- s1a decide asks the chosen System One model for the next action.
- The agent executes the action and repeats until done; runs are logged.
Strengths
- Same agents run on Jev, open models or a chat model for comparison.
- Reported speed and cost gains over a chat model on the shipped scenarios.
- Claude Code and Codex integration via plugin and MCP.
Caveats
- Benchmark numbers come from the authors' own runs; one chat-model cost is estimated.
- Jev runs need a TypeSafe key.
- Initial setup resolves a pinned dependency and takes several minutes.
Getting started
uv sync && cp .env.example .env, then uv run s1a run blackjack --model jev --rethink off --episodes 20.At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- CLI + Claude Code plugin + MCP server
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
System 1 decision models (Jev, Laya, Cua-S1) as brain for agents: Browser use…
wh000wh000/awesome-jev-livePrebuilt agents whose decisions come from a System One model, Jev or an open one, covering browser use, computer use, robotics and games, with a scaffold for building your own.
MrJev/awesome-jev
Frequently asked questions
What does the System One model decide?
The next action at each agent step, such as which element to click, which game move to make or which queue a ticket goes to.
Which models are supported?
Jev, Laya and Cua-S1, plus OmniJev for browser agents and a chat model for comparison runs.
Can I use it from Claude Code?
Yes, install the s1a plugin from the system1-agents marketplace to run or build agents.
Related projects
Lets coding agents run whole goals in your signed-in Chrome on macOS in background tabs, asking Jev or Claude Haiku only for real choices.
ProjectbetaCodex plugin adding a bounded Jev decision loop to Computer Use for browser tabs and native macOS apps.
ProjectbetaWindows and Chrome GUI delegation for Codex: a local controller runs a task contract, calling Jev only for semantic choices.
ProjectprototypePythonComputer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonLocal stdio MCP server that executes whole browser and macOS goals with a Jev-driven observe-decide-act loop.
MCP serverbetaTypeScriptBrowser Use's browser agent where Jev picks an operation and page element in one request and a small LLM only writes text to type.
LibrarybetaPython