Harness

Mac app that sends an AI agent through your app's UI with a goal and a persona, then reports whether it succeeded and where it hit friction. Includes an MCP server.

View on GitHub ↗
Stars
★ 357
Last commit
2 Sept 2026
Maintenance
Active
Licence
MIT
Category
MCP Servers
GitHub
awizemann/harness

Overview

Harness is a native macOS developer tool that uses an AI agent to run user tests against an iOS Simulator app, a macOS app or a web app. You give it a goal in plain language and a persona, and the agent works through your app from screenshots, notes any UX friction and stops when it succeeds, fails or gives up. It also has an MCP server so other agents can drive it.

Features

  • Reports whether the goal was completed, the path taken as a replayable sequence, and timestamped friction points
  • Drives iOS Simulator builds through xcodebuild, simctl and WebDriverAgent
  • Drives macOS apps through accessibility actions without moving your real pointer
  • Drives web apps in an embedded WKWebView at desktop or mobile size
  • MCP server for autonomous runs or step-by-step UI sessions with no API key needed
  • Works with cloud models from Anthropic, OpenAI and Google, or local models through Ollama

Requirements

  • macOS 14 or later (Apple silicon or Intel)
  • Screen Recording and Accessibility permissions for macOS targets
  • An API key for a cloud model provider for autonomous runs, or Ollama for local models

Install

  1. Download the universal macOS build from the GitHub releases page

Example

Give Harness the goal "I want to sign up and create my first list" with the persona "first-time user, never seen this app" and review where the agent got stuck.

Works with

Claude CodeAny MCP client

Listed in

swiftskills.dev

Read the code before you install anything. SwiftAgents doesn't review or endorse the tools listed here. Active: a commit in the last 90 days. Slowing: last commit within the past year. Stale: older than a year, or archived.

More in this category