System1-Agents plugin
Skill, plugin and MCP server: fast Jev and Laya models make choices instead of a pricey chat model in browsers, games and routing.
- 182 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Agents click through pages and drive the computer, and a browser run opens real websites.
- Task descriptions and state are sent to the model service under your key.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/system1-agentsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Agents click through pages and drive the computer, and a browser run opens real websites.
- Task descriptions and state are sent to the model service under your key.
This is third-party code. Review the repository files before installing.
What it does
System1-Agents gives an agent a fast way to make decisions. A coding agent identifies a task such as routing a ticket, picking a page element or choosing a game move and, instead of calling a large chat model, runs `s1a decide`. It returns the chosen option, a probability for each and the response time in milliseconds. There are three ways to connect. The skill installs through `npx skills add` for Claude Code, Codex and Cursor. The Claude Code plugin adds a browser subagent and an MCP server. A standalone `s1a-mcp` server exposes `list_agents`, `run_agent` and `decide` to any host. Prebuilt agents cover browser use, computer use, robotics and games. According to the README table, answers come up to six times faster than a chat model and up to twenty-five times cheaper on the same tasks.
Who it is for. Teams that run agents through repetitive choices such as support ticket routing and track the cost of every model call.
Good fit when
- Support tickets must be sorted into five queues within hundreds of milliseconds.
- An agent plays a game or clicks through a page and a chat model call at every step costs too much.
- You want your own decision agent on Jev, Laya or Cua-S1 Nano.
- A coding agent should hand off one-off option selections without loading the main model.
Not a fit when
- The task needs arithmetic, deduction or free text: those stay with a chat model.
- You have no Jev (`TYPESAFE_API_KEY`) or OpenRouter key and do not want a local Cua model.
- You want a simple prompt skill without Python and `uv`.
Example request
Route this ticket to logistics, payment, returns, account or human: "I was charged twice for order 4411 and I want the second charge refunded."Limitations
Command-line use needs `uv` and `uv sync`; the first sync resolves the pinned openjiuwen version and takes a few minutes. Jev models require a `TYPESAFE_API_KEY` from the TypeSafe console or `OPENROUTER_API_KEY`. The Cua model answers locally after `uv sync --extra cua`. A browser agent run lasts minutes, so the host tool timeout must go above the 60-second Codex default. Speed and price figures come from the authors benchmarks on a few scenarios, so your task will differ.
How to disable. Remove the skill through your skills manager and the plugin with `claude plugin uninstall s1a`. Drop the MCP server with `codex mcp remove s1a` or from the host config.
FAQ
What does `s1a decide` return?
One JSON object: the chosen `choice`, a probability per option, a `confidence` field and `ms`. Options are set with the repeatable `--option KEY=DESCRIPTION` flag.
How do I add the MCP server in Codex?
With `codex mcp add s1a -- uv run --project /path/to/system1-agents s1a-mcp`. Run `uv sync` in the clone first.
Which tools does the MCP expose?
Three: `list_agents`, `run_agent` and `decide`.
Which models are supported?
Jev from TypeSafe, Laya and Cua-S1 Nano. For comparison the same agent also runs on a chat model with `--model llm`.
Can I build my own agent?
Yes, the repository includes a `build-s1a-agent` skill and a Build your own section. It leads from a task description to a new agent.
Discussion
No questions or reviews yet.
Related
PluginsA plugin of seven skills for Claude Code and Codex: prompts, subagents, skill authoring, editing, web search and a browser.
Self-hosted system for a team of AI agents: org chart, tasks, budgets and oversight of Claude Code, Codex and OpenClaw
Agent-Native Agent
Agent-Native Agent (ANA)
A Claude Code plugin and self-hosted app: a dashboard and chat with a coding agent that rewrites the app while you use it
Anthropic skill for Claude Code that audits a multi-agent system and, by measurement, folds subagents into one loop with skills.
A skill registry for teams with security scanning and versioning; the MCP server finds and installs skills on request.
Claude Code plugin with 60 subagents and 62 team presets, plus a browser designer for planning agent teams visually.
README in short
The repository promises agents a System 1 decision model: fast one-step choices instead of chat model reasoning. A table compares Jev with a chat model on time and price across seven scenarios. Instructions for the skill, the browser subagent and building your own agent sit alongside.