# Agent Pit Stop

> An open source design system for agent interfaces: UX principles with live
> demos and production React components for the moments where agents and
> humans sync (approval, inspection, handoff, recovery). MIT licensed.
> Components depend only on React, Tailwind, and Motion, and install via the
> shadcn registry. Site: https://agent-pitstop.vercel.app

## Install

Any component: npx shadcn@latest add https://agent-pitstop.vercel.app/r/<name>.json
Catalog (JSON): https://agent-pitstop.vercel.app/r/index.json
Dense contract for agents (props, principles, guarantees): https://agent-pitstop.vercel.app/r/contract.json
This file, the catalog, and the contract are generated from component source at build time; they cannot drift.

## Components

- [AgentInbox](https://agent-pitstop.vercel.app/components/agent-inbox): The review queue for background agents. (source: https://agent-pitstop.vercel.app/r/agent-inbox.json)
- [AgentRoster](https://agent-pitstop.vercel.app/components/agent-roster): The team sheet for multi-agent runs: who's working, who's waiting, who's blocked, and on what. (source: https://agent-pitstop.vercel.app/r/agent-roster.json)
- [AgentTaskList](https://agent-pitstop.vercel.app/components/agent-task-list): The agent's plan, made visible: what it will do, what it's doing, what's done, and what broke. (source: https://agent-pitstop.vercel.app/r/agent-task-list.json)
- [ApprovalGate](https://agent-pitstop.vercel.app/components/approval-gate): An inline permission request: the agent states what it wants to do, shows the exact scope, and waits. (source: https://agent-pitstop.vercel.app/r/approval-gate.json)
- [ConfidenceMeter](https://agent-pitstop.vercel.app/components/confidence-meter): Honest uncertainty display: confidence as coarse bands, never decimal theater, paired with a verification affordance exactly when doubt is warranted. (source: https://agent-pitstop.vercel.app/r/confidence-meter.json)
- [ContextBudget](https://agent-pitstop.vercel.app/components/context-budget): A delegation contract's limits, made visible: tokens, spend, or time against their caps. (source: https://agent-pitstop.vercel.app/r/context-budget.json)
- [InterruptBar](https://agent-pitstop.vercel.app/components/interrupt-bar): The steering wheel and the brake for a running agent. (source: https://agent-pitstop.vercel.app/r/interrupt-bar.json)
- [ToolCallCard](https://agent-pitstop.vercel.app/components/tool-call-card): Shows an agent's tool call with legible, progressive disclosure. (source: https://agent-pitstop.vercel.app/r/tool-call-card.json)

## Principles

- [Legible Thinking](https://agent-pitstop.vercel.app/principles/legible-thinking): agents lose trust by being wrong invisibly; show intent, state, and evidence
- [Interruptibility](https://agent-pitstop.vercel.app/principles/interruptibility): users delegate more to agents they can stop; steer beats stop
- [Delegation Contracts](https://agent-pitstop.vercel.app/principles/delegation-contracts): itemize scope before consent; boundaries beat permissions
- [Calibrated Trust](https://agent-pitstop.vercel.app/principles/calibrated-trust): bands not decimals; verification cheaper than redoing
- [Graceful Failure](https://agent-pitstop.vercel.app/principles/graceful-failure): failure is a rendered state; partial work is never discarded

## Docs

- [How it works](https://agent-pitstop.vercel.app/how-it-works)
- [Rules for the pit](https://agent-pitstop.vercel.app/principles)
- [Contributing](https://github.com/acaspx/agent-pitstop/blob/main/CONTRIBUTING.md)