harness

Harness

macOS 14+ · Apple Silicon & Intel · ~12 MB

awizemann.github.io/harness  ·  Wiki  ·  All releases

A native macOS developer tool that drives an iOS Simulator, a macOS app, or a web app with an AI agent so you can run user tests — not scripted UI tests, but real-user simulation.

You write a goal in plain language ("I want to sign up and create my first list", "delete my account", "find a vegetarian restaurant near me and save it") and a persona ("first-time user, never seen this app"). Harness builds (or just launches) your target, and an LLM agent reads screenshots, clicks/types/scrolls, and pursues the goal — narrating what it sees, flagging UX friction (dead ends, ambiguous labels, unresponsive controls), and stopping when it succeeds, fails, or would give up.

Three artifacts come out of every run:

1. Did the goal complete? — success / failure / blocked + summary
2. What was the path? — replayable sequence of screens + actions
3. Where was the friction? — timestamped events the agent flagged as confusing

Targets

Per-app setting: each Application declares its kind once at create time. The agent's tool schema (clicks vs swipes vs key shortcuts vs navigate) and the system-prompt context block re-shape per platform. Run history, replay, and friction reporting are platform-neutral.

Status: v0.8.1 (alpha). Drive Harness from an agent via the MCP server — either autonomous runs (surfaced as first-class, badged history) or step-level UI sessions that let an external client see and act on a web, iOS, or macOS target directly, with no LLM loop and no API key. The harness-mcp binary is now standalone and relocatable — bundle it inside another product and run it from anywhere. Sparkle auto-update built in. All three platforms wired end-to-end with Set-of-Mark targeting on iOS, macOS, and web (numbered overlays on interactive elements; agent clicks by id, not pixel; agent-only, never on disk); Local Mac inference via Ollama (Qwen3-VL 8B, Gemma 4 Vision 9B, Llama 3.2 Vision 11B, plus a custom-model field) alongside cloud providers (Anthropic Opus 4.7 / Sonnet 4.6 / Haiku 4.5; OpenAI GPT-5 Mini / GPT-4.1 Nano; Google Gemini 2.5 Flash / Flash Lite); per-Application credential storage; per-provider Keychain storage; configurable per-model token budgets; unlimited-step option; harness-cli dev-time driver. macOS needs Screen Recording + Accessibility permission. Web is WebKit-only; Chrome via CDP is on the roadmap. See docs/ROADMAP.md.

What's new in 0.8.1

• Authenticated web sessions — inject and export session state. Web sessions stay fresh-user by default, but SSO-only products were unreachable. start_ui_session gains web-only visible: true (the session window comes on screen for a human to log in — password, SSO, MFA) and session_state (cookies + localStorage, injected before the first navigation). The new export_ui_session_state tool returns the live session's state in exactly the shape session_state accepts, so a human-authenticated visible session round-trips into later headless ones. Cookie values follow the fill_credential precedent: never logged, never in steps.jsonl, never on disk, never on the persisted run model. See HarnessMCP/README.md.
• Mark labels come from the label, not the placeholder. The web probe now resolves an accessible name the way a screen reader would (aria-label → labelledby → <label> → placeholder → title → value → text → name) and each structured mark reports which rule won as label_source (declared in the outputSchema; web only). No more selectors keyed on sample placeholder data.
• act_ui settles on the DOM change, not just on navigation. Observation is armed before dispatch and the gate drains the page's own pending work (setTimeout ≤ 2s, in-flight fetch/XHR), so a same-URL React state swap lands in the action's own observation instead of the pre-action frame. Idle pages still return at the 250ms floor; ceiling 3s.
• 359 unit tests passing (326 at 0.8.0), plus three new live-smoke phases: label priority, the 500ms same-URL swap captured by act_ui itself, and the injection → export round trip with a steps.jsonl secret check.

Notes for earlier versions live on the Releases page.

First clone

Harness vendors appium/WebDriverAgent as a git submodule under vendor/WebDriverAgent (it's how we drive the iOS Simulator's responder chain). The Xcode project is generated from project.yml via xcodegen.

git clone https://github.com/awizemann/harness.git
cd harness
git submodule update --init --recursive
brew install xcodegen
xcodegen generate
open Harness.xcodeproj

The first run builds WDA against your simulator's iOS runtime (~1–2 min). Result is cached under ~/Library/Application Support/Harness/wda-build/<iOS-version>/ and reused on subsequent runs.

Full setup: see Build-and-Run on the Wiki.

How to read this repo

• standards/INDEX.md — development, code, and architecture standards. Read these before adding code.
• GitHub Wiki — "where things live, why, and how to extend them." Maintained per PR alongside code.
• docs/ARCHITECTURE.md — system architecture overview.
• docs/ROADMAP.md — build order and milestones.
• docs/PROMPTS/ — canonical agent prompts (loaded as a bundle resource at runtime).
• HarnessDesign/ — design system tokens, primitives, and screen layouts.

Contributing

PRs welcome. Read CONTRIBUTING.md first — it covers setup, the architecture rules (MVVM-F, Swift 6 strict concurrency, single subprocess actor), and the public-surfaces sync rule (code changes that affect README / wiki / site update them in the same PR).

License

MIT — see LICENSE.