Describe a step in plain English. AppClaw figures out what to tap, type, and swipe on real Android & iOS devices, from the CLI or the Cursor & VS Code extension.
Describe a high-level goal in plain English. AppClaw breaks it into sub-goals, runs a perceive–reason–act loop on each, and verifies the screen before it ever says “done”, recovering on its own when the screen fights back.
AppClaw works from two context sources, the knowledge you hand it, and the experience it earns. Both feed the same context window, so the agent gets the right detail exactly when it needs it.
Per-app navigation maps injected into the agent's context, where the compose button lives, which gesture archives, how to reach any setting. No trial-and-error exploration.
After every successful run, AppClaw records the winning steps. On the next run it retrieves the most relevant past trajectories, matched by screen and goal, and injects them as a compact hint.
Two halves of one job. AppClaw reads what's on the screen, then pinpoints exactly which element you mean, no brittle selectors.
A “Login” button and a “Log in” text link on the same screen? Skip brittle selectors. Reference the one you mean by its position relative to another element, AppClaw resolves it geometrically.
It reads the structured accessibility tree when it's there, fast, exact, cheap. When it isn't (Canvas, Flutter, games, custom-rendered UI), it falls back to AI vision on the raw screenshot. The same tap, either way.
Add one line and AppClaw fans your suite across the whole rack, each device with its own screen, step tracking, and pass/fail.
An AI agent for messy goals, deterministic YAML for repeatable tests, or a live playground to build flows by hand.
Describe a goal in plain English. AppClaw reads the screen, reasons, and executes until it's finished.
Write plain-English YAML. No AI needed when you run it, fast, repeatable, with vision fallback.
A REPL on a real device. Type commands, watch them run, export straight to YAML.
Plain English as a prompt, a YAML file, or a playground command. No selectors.
AppClaw reads the screen, sends it to your LLM, and gets back a structured plan. Adapts to popups.
Actions run through Appium via MCP. It checks after every step until the goal is met.
Every device action is delegated over the open MCP protocol, no drivers, no vendor lock-in.
Reads the screen, decides the next move, and verifies the result, looping until the goal is met. That's all it does.
From a prompt to a passing test, AppClaw drives the whole flow, plus a playground and per-app AppGuides for the steps that need a hand.
How the agent reads a screen, when to reach for YAML, and where AI is taking mobile testing.
The honest worry: if an LLM drives every step, every run costs money. It doesn't have to. Here's what actually happens behind each step, why the model in DOM mode only parses language and never sees the screen, and how AppClaw caches and heals locators so repeat runs skip it entirely.
AppClaw reads the structured accessibility tree when it exists (fast, exact, cheap) and falls back to vision on the raw pixels when it doesn't. Here's how the agent decides, and why the same tap works either way.
Selectors describe how an app is built. Instructions describe what you want. AppClaw closes the gap by resolving intent on the live screen instead of matching a fixed path.
Open source, extensible, BYO LLM key. Start in under a minute, terminal or IDE.