System1 Agents MCP Server

<h1 align="center" id="system1-agents"> <picture> <source media="(prefers-color-scheme: dark)" srcset="docs/assets/logos/system1-agents-dark.png"> <img src="docs/assets/logos/system1-agents.png" alt="System1-Agents" width="760"> </picture> </h1>

[!NOTE] Give your agents a System 1 decision model. Start from a prebuilt agent or build your own.

Describe the task. Claude Code or Codex runs a prebuilt agent or builds a new one, on Jev, Laya or Cua-S1 Nano. Browser use, computer use, robotics and games ship ready to run.

Up to 6× faster and 25× cheaper than a chat model, at the same score.

Quickstart · From Claude Code or Codex · Benchmarks · Docs

<!-- placeholder: name the chat model behind the "chat model" columns (the MODEL_NAME of these runs). -->

scenario	Jev	chat model	speedup	Jev cost	chat model cost	more expensive
Browser use (Allrecipes)†	35.7 s	138.7 s	3.89×	$0.091356	$1.494331	16.4×
Custom agent (ticket router, 30 tickets)	12.7 s	65.3 s	5.14×	$0.000764	$0.014800	19.4×
Computer use (Windows Calculator)	17.2 s	30.1 s	1.75×	$0.000447	$0.003002	6.7×
Robotics (ALFWorld)	7.4 s	25.7 s	3.47×	$0.000221	$0.002800*	12.7×
Games (2048, 20 moves)	27.1 s	68.5 s	2.53×	$0.000447	$0.006139	13.7×
Games (Millionaire)	15.3 s	21.6 s	1.41×	$0.000168	$0.000892	5.3×
Games (Blackjack)	2.3 s	14.7 s	6.39×	$0.000021	$0.000531	25.3×

† The first Allrecipes task of the WebVoyager task set (He et al., 2024, Apache-2.0, attribution in NOTICE): a vegetarian lasagna with over 100 reviews, 4.5 stars or more, for 6. The chat model of that row is Claude Fable 5.1 through OpenRouter; both models pay it for the typed search text and the answer. * Estimated; the chat-model run recorded no cost. Each replay below is the episode behind its row, Jev on the left and the chat model on the right, both on the wall clock. The other Allrecipes runs, longer games and the Google Flights driver comparison with its rerun of 2026-09-23: docs/benchmarks.md.

<table> <tr> <td width="50%"><img src="docs/assets/demos/allrecipes-comparison-8x.gif" alt="Allrecipes: a vegetarian lasagna search on the live site, Jev on the left, the chat model on the right, with Jev's probabilities over the page's controls under each step"><br>_{Browser use, WebVoyager's Allrecipes task 0 on the live site, replay at 8× speed}</td> <td width="50%"><img src="docs/assets/demos/desktop-comparison-4x.gif" alt="Windows Calculator: clicks toward 12 times 7 until the display shows 84, Jev on the left, the chat model on the right"><br>_{Computer use, the Windows Calculator, replay at 4× speed}</td> </tr> <tr> <td><img src="docs/assets/demos/ticket-router-comparison-8x.gif" alt="Ticket router: 30 labelled tickets routed to five queues, Jev on the left, the chat model on the right"><br>_{Ticket router, 30 tickets to five queues, replay at 8× speed}</td> <td><img src="docs/assets/demos/alfworld-validated-comparison-2x.gif" alt="ALFWorld household task with its AI2-THOR scene, Jev on the left, the chat model on the right"><br>_{ALFWorld, an embodied household task, replay at 2× speed}</td> </tr> <tr> <td><img src="docs/assets/demos/2048-comparison-8x.gif" alt="2048 over 20 moves, Jev on the left, the chat model on the right"><br>_{2048, 20 moves, replay at 8× speed}</td> <td><img src="docs/assets/demos/millionaire-comparison-4x.gif" alt="Millionaire quiz ladder, Jev on the left, the chat model on the right"><br>_{Millionaire, a 15-question quiz ladder, replay at 4× speed}</td> </tr> <tr> <td><img src="docs/assets/demos/blackjack-comparison-2x.gif" alt="One Blackjack hand, Jev on the left, the chat model on the right"><br>_{Blackjack, one hand, replay at 2× speed}</td> <td></td> </tr> </table>

OmniJev: a decision model that reads the screen

OmniJev (Apache-2.0, Beijing Zhongguancun Academy, CASIA and Zevo) is a System 1 decision model on Qwen3.5 vision-language backbones (0.8B, 2B, 4B): it answers the same typed questions as Jev over a screenshot, a video or a robot camera. --model omnijev puts it in the slot of the browser agents, which then send it a screenshot of the page at every step (docs/decision-models.md).

The clips below are OmniJev's own v1.1 demos with its 4B model: replays of recorded trajectories with the model's probabilities, not s1a runs and not live control. They are shown from the OmniJev repository and keep their upstream terms.

<table> <tr> <td width="50%"><img src="https://raw.githubusercontent.com/tinnel123666888/OmniJev/14dbec4f71e194852c8d7b88ab36ef639493f400/docs/media/v11/web.gif" alt="OmniJev on Mind2Web web tasks: a Central Park to JFK route and a nightstand comparison, with the model's probabilities per step"><br>_{Web: Mind2Web test tasks (OmniJev v1.1 replay)}</td> <td width="50%"><img src="https://raw.githubusercontent.com/tinnel123666888/OmniJev/14dbec4f71e194852c8d7b88ab36ef639493f400/docs/media/v11/phone.gif" alt="OmniJev on AndroidControl phone tasks: London weather, a 59-minute timer and a drawing tutorial"><br>_{Phone: AndroidControl tasks (OmniJev v1.1 replay)}</td> </tr> <tr> <td><img src="https://raw.githubusercontent.com/tinnel123666888/OmniJev/14dbec4f71e194852c8d7b88ab36ef639493f400/docs/media/v11/arcade.gif" alt="OmniJev on Atari recordings: Enduro, Skiing and Pong, next input and steering"><br>_{Games: Enduro, Skiing and Pong (OmniJev v1.1 replay)}</td> <td><img src="https://raw.githubusercontent.com/tinnel123666888/OmniJev/14dbec4f71e194852c8d7b88ab36ef639493f400/docs/media/v11/robot.gif" alt="OmniJev on a two-view Bridge robot recording folding a cloth: jog direction, gripper and move size"><br>_{Robotics: folding a cloth, two camera views (OmniJev v1.1 replay)}</td> </tr> </table>

Choose your path

From Claude Code or Codex

npx skills add ThinkFlowLab/system1-agents                                                              # the skill: Claude Code, Codex, Cursor
claude plugin marketplace add ThinkFlowLab/system1-agents && claude plugin install s1a@system1-agents   # plus the browser subagent and the MCP server

The skill tells the host when to hand a task to a System 1 agent and which command to run. The ticket router from the table above, from Claude Code:

Route this ticket to logistics, payment, returns, account or human: "I was charged twice for order 4411 and I want the second charge refunded."

The host runs one s1a decide over the five queues and reports the queue with its probability, in about 400 ms. Page tasks, games and one-off selections go the same way; arithmetic, deduction and free text stay with the chat model. The walkthrough, the plugin's keys and what to delegate: docs/skills.md.

From the command line

git clone https://github.com/ThinkFlowLab/system1-agents && cd system1-agents
uv sync && cp .env.example .env     # the first sync resolves the openjiuwen pin and takes a few minutes

Put a Jev key in .env (TYPESAFE_API_KEY from the TypeSafe console, or OPENROUTER_API_KEY), then ask for one decision and run one agent with each model:

uv run s1a decide --state '{"player_total": 18, "dealer_upcard": 9}' \
  --option hit="take a card" --option stand="keep the hand" --rules "stand on 17 or more"
uv sync --extra blackjack
uv run s1a run blackjack --model jev --rethink off --episodes 20
uv run s1a run blackjack --model llm --rethink off --episodes 20     # the chat model in the same agent

Each --option KEY=DESCRIPTION must use a unique, case-sensitive key. Repeated keys are usage errors (exit 2), reported before loading a model or checking its credentials; they never replace an earlier description.

decide prints one JSON object with choice, a probability per option, confidence and ms; run writes a job folder with the score. Without a key, --model cua answers in process after uv sync --extra cua.

As an MCP server

codex mcp add s1a -- uv run --project /path/to/system1-agents s1a-mcp

It serves three tools, list_agents, run_agent and decide, to any host that speaks MCP. Run uv sync in the checkout once before registering it, and raise the host's tool timeout: a browser run takes minutes, beyond Codex's 60 s default. The settings for Codex and Claude Code: docs/skills.md.

Build your own

A System 1 agent is one module under s1a/agents/ that ends in a frozen SPEC; Blackjack is 111 lines. The builder skill runs in Claude Code from this checkout, probes the task with 8 to 12 hand-written decisions before it writes code, and stops when the task needs deduction or arithmetic:

Build a System 1 agent for <task>.

The gates and the templates: docs/skills.md.

What ships

• allrecipes: browser use, WebVoyager's first Allrecipes task, headed on the live site.
• flights: browser use, a Google Flights search over @playwright/mcp.
• desktop: computer use, any Windows or macOS app window through Cua Driver.
• ticket_router: 30 labelled support tickets to five queues.
• alfworld: household tasks in text, with the AI2-THOR scene in the replays.
• game2048, millionaire, blackjack: games with a score per episode.
• injection_guard: a rail that answers one question at a hook of a running agent and fails closed.

These use cases are application candidates for the required application + System1-Agents + System1-Omni video in important PRs.

Every agent runs on jev, laya or cua, and on the chat model for the comparison; the browser agents also run on omnijev, which decides over a screenshot. Flags, run commands and extras: docs/agents.md.

How it works

Each agent is a stock openJiuwen agent with a System 1 decision model as its model. On a decision turn the model gets the state and the options and answers with one of them; planning, typed values and the final answer stay with the chat model in the same agent. Any decision model with that interface fits: docs/architecture.md, docs/decision-models.md.

Docs

• docs/benchmarks.md: the six runs above, the Google Flights driver comparison and its 2026-09-23 rerun, a longer game, the guard rail.
• docs/skills.md: the caller skill, the builder skill, what to delegate.
• docs/agents.md: every agent with its flags, run command and extra.
• Agent use-case recipes: setup, run commands, result checks and demos for complete tasks.
• docs/architecture.md and docs/decision-models.md: the fronts, the model slot, the model interface, adding a backend.
• docs/browser-front.md: the browser policy, decision by decision.
• docs/served-laya.md: Laya served by system1-omni as a decision model over HTTP, with its API spec.
• docs/clm.md: CLM as a decision model over HTTP (--model clm), the engine that owns everything after a frozen Qwen3-8B encoder.
• docs/configuration.md: environment variables, defaults and reader subsystems in one table.
• docs/glossary.md: terms the documentation glosses on first mention.
• docs/why.md: the problem, the philosophy, the precedents.
• docs/roadmap.md: application scenarios, benchmarks, model backends and System1-Omni integration; CHANGELOG.md.

Stay Tuned with Us

If you find system1-agents useful, give us a star on GitHub to support the project and help others discover it!

Contributing and license

CONTRIBUTING.md has the dev install, the checks and the hooks. Apache-2.0.