OverviewHistoryStatsSecurity
npx skills add ...
Documentation
SKILL.md
npx skills add web-infra-dev/midscene-skills --skill browser-automation
Vision-driven browser automation using Midscene. Operates from screenshots — no DOM or accessibility labels needed.
npx skills add web-infra-dev/midscene-skills --skill browser-automation
CRITICAL RULES — VIOLATIONS WILL BREAK THE WORKFLOW:
- Never run midscene commands in the background. Each command must run synchronously so you can read its output (especially screenshots) before deciding the next action. Background execution breaks the screenshot-analyze-act loop.
- Run only one midscene command at a time. Wait for the previous command to finish, read the screenshot, then decide the next action. Never chain multiple commands together.
- Allow enough time for each command to complete. Midscene commands involve AI inference and screen interaction, which can take longer than typical shell commands. A typical command needs about 1 minute; complex
actcommands may need even longer.- Always report task results before finishing. After completing the automation task, you MUST proactively summarize the results to the user — including key data found, actions completed, screenshots taken, and any relevant findings. Never silently end after the last automation step; the user expects a complete response in a single interaction.
Automate web browsing using npx -y @midscene/web@1. By default, launches a headless Chrome via Puppeteer that persists across CLI calls — no session loss between commands. Also supports CDP mode and Bridge mode to connect to an existing Chrome browser.
act Can DoInside a single act call in the browser, Midscene can click, right-click, double-click, hover, type or clear text, press keys, scroll, drag, long-press, and continue through multi-step page flows based on what is currently visible. When touch input is enabled, it can also handle swipe- or pinch-style interactions on touch-oriented pages.
This skill has three modes. Choose based on the user's intent:
| Mode | When to use | How it works |
|---|---|---|
| Puppeteer (default) | User wants to browse a URL, scrape data, test UI — no need for their own browser | Launches a new headless Chrome, isolated from user's browser |
| CDP mode | User says "connect to my Chrome", "control my browser", "CDP", "remote debugging", or wants to operate their existing browser. Also use when the task implicitly requires login state (e.g., "check my orders", "open my dashboard", "look at my account") | Connects to user's Chrome via DevTools Protocol. Requires remote debugging enabled (chrome://inspect > "Allow remote debugging"). No extension needed |
| Bridge mode | User explicitly mentions "bridge", "extension", or has Midscene Chrome Extension installed and prefers to use it | Connects to user's Chrome via the Midscene Chrome Extension |
CDP vs Bridge: Both control the user's real Chrome with login sessions preserved. CDP only needs a Chrome setting toggle; Bridge needs a Chrome Extension installed. If the user doesn't specify, prefer CDP mode as it has fewer prerequisites.
Before using CDP mode, run a quick precheck to verify Chrome's remote debugging port is reachable. This avoids long timeouts when the user hasn't enabled remote debugging.
How to use the precheck result:
101 → CDP mode is available, use --cdp000) → Chrome is not running with remote debugging enabled. Start Chrome with --remote-debugging-port=9222, wait 2-3 seconds, then re-run the precheck. If it still fails, fall back to Puppeteer mode (or use Bridge mode if the Midscene Chrome Extension is installed).Bridge mode has no usable precheck. Despite port 3766 being involved, the bridge server is started by the CLI (
npx ... --bridge connectopens 3766 itself); the extension is the client that hooks in. Checking 3766 before the CLI runs always returns nothing. Just run--bridge connectdirectly — its first log line (waiting for bridge to connect...→one client connected) tells you whether the extension picked it up.
Midscene requires models with strong visual grounding capabilities. The following environment variables must be configured — either as system environment variables or in a .env file in the current working directory (Midscene loads .env automatically):
Example: Gemini (Gemini-3-Flash)
Example: Qwen 3.5
Example: Doubao Seed 2.0 Lite
Commonly used models: Doubao Seed 2.0 Lite, Qwen 3.5, Zhipu GLM-4.6V, Gemini-3-Pro, Gemini-3-Flash.
If the model is not configured, ask the user to set it up. See Model Configuration for supported providers.
Use CDP mode to control the user's existing Chrome browser. The default CDP endpoint is ws://127.0.0.1:9222/devtools/browser (port 9222 is Chrome's standard remote debugging port). If the user specifies a different port, replace 9222 accordingly.
Add --cdp <ws-endpoint> to every command:
When the user needs custom request headers in CDP mode, such as routing requests to a PPE environment, pass each header through a separate --extra-http-header 'Name:Value' option. Repeat the option to send multiple headers. Use the exact header names and values supplied by the user; do not guess environment names or authentication values.
Each CLI command creates a new CDP session. Repeat every required --extra-http-header 'Name:Value' option on each CDP command that may issue requests, including connect, act, and assert. Split each entry at the first colon, so values may contain additional colons. The headers are applied before connect --url navigates, so the initial document request includes them. Do not print header values in task summaries, and avoid putting sensitive authentication values directly in shell history.
disconnect releases the connection but does NOT close the browser. There is no close command in CDP mode.connect --url navigates the existing active tab instead of opening a new tab.connect without --url attaches to the current active tab without navigating.chrome://inspect in Chrome and turn on "Allow remote debugging".Use Bridge mode when the user explicitly mentions "bridge", "extension", or has the Midscene Chrome Extension installed. Add --bridge to every command:
127.0.0.1:3766); the extension is the client that connects to it. The "Listening" status shown in the extension panel means "ready to connect to a CLI", not that the extension itself opened a port. Practically: launch --bridge connect first; the extension will hook in within a second or two.chrome://, chrome-extension://, the Chrome Web Store, or other privileged URLs. If the active tab is one of those, the CLI will return Cannot access a chrome:// URL. Ask the user to switch to a regular http(s):// tab before reconnecting.chrome://extensions, find Midscene, click "Details", enable the switch, then switch back to the target http(s):// tab before reconnecting. If upload fails with a permission or Not allowed error, check this first.connect without --url attaches to the current active tab; connect --url <href> navigates that tab. The CLI does not open new tabs in bridge mode.disconnect only closes the CLI-side bridge connection, not the browser or tabs.After taking a screenshot, read the saved image file to understand the current page state before deciding the next action.
Use act to interact with the page and get the result. It autonomously handles all UI interactions internally — clicking, typing, scrolling, hovering, waiting, and navigating — so you should give it complex, high-level tasks as a whole rather than breaking them into small steps. Describe what you want to do and the desired effect in natural language:
When a prompt asks Midscene to upload files, act requires an explicit --file-chooser-allowed-dir. Use the smallest directory that contains the test fixtures; do not grant the project root or a home directory. Refer to files by paths relative to that directory in the prompt.
In Bridge mode, also confirm the Midscene extension has "Allow access to file URLs" enabled in chrome://extensions > Midscene > "Details", and reconnect from the target http(s):// tab after changing the switch.
Use assert to verify that the current page satisfies a natural language condition. It does not perform UI actions; it checks the visible page state and passes only when the assertion is true. Use this for validation, QA checks, and final state verification after act.
In CDP or Bridge mode, pass the same connection flags you use for other commands:
By default a failed assertion throws an AI-generated reason. Pass --message to throw a custom error message instead, which is useful for surfacing the intended outcome in QA and CI logs.
When the assertion needs to compare against a reference image (icon, logo, screenshot), pass --image for the URL/path and --image-name for its display name. Each --image may be an http(s) link, a data: URI, or a local file path. Repeat both flags in matching order when you need to attach more than one image. Add --convertHttpImage2Base64 true when the model cannot reach the URL directly. Requires @midscene/web@1.9.0+.
Use a recording when the state to verify may disappear before a current-screen assertion runs, such as a toast, loading banner, animation, or transition. Run the recorder in a dedicated interactive terminal:
Pass the normal target flags and --output to record start, then wait for Recording. Press Ctrl+C to stop and save. Keep the recorder as a foreground process in its terminal; never add shell &. Perform the interaction manually or from a second terminal, send Ctrl+C to the recorder, and wait until it prints the saved path before asserting. In CDP mode, repeat --cdp ws://127.0.0.1:9222/devtools/browser on record start, actions, and assert. In Bridge mode, the foreground recorder owns the single Bridge connection, so interact manually instead of starting a second Bridge CLI action.
Optional capture flags on record start are --interval-ms, --max-frames, and --watchdog-ms; --max-frames caps sampled frames, and the manifest may contain one additional final representative frame. The default watchdog finalizes and saves the recording after five minutes, while --watchdog-ms 0 disables that safety limit. The output is a JSON manifest plus an adjacent <name>.frames image directory, not an encoded video or archive. The manifest contains relative JPEG/PNG paths and no base64 image bodies. Keep or move the JSON file and image directory together, and pass the JSON path to assert --record. Use ordinary assert without --record when only the current page matters.
When the user provides a screenshot, icon, logo, or reference image and wants an exact visual match, prefer tap --locate instead of a generic act --prompt. Pass --locate as JSON. The prompt describes the target, images supplies named reference images, and convertHttpImage2Base64: true is useful when the image URL may not be directly accessible to the model.
The same locate JSON shape also works for other commands that accept a locate parameter.
Disconnect from the page but keep the browser running:
Close the browser completely when finished (Puppeteer mode only):
The generated HTML report is recommended for human reading first. It includes step-by-step execution details and replay videos for each operation, which makes it much easier to understand what happened and troubleshoot problems.
If another skill or tool needs to consume the report, first convert it with report-tool from the same platform CLI package. Prefer Markdown for LLM-based workflows. Use JSON when the report needs to be processed programmatically.
The browser persists across CLI calls via a background Chrome process. Follow this pattern:
act to perform the desired action or target-driven instructions. Use assert for the resulting page state, or keep record start --output ... running in a dedicated terminal during transient-state workflows, stop it with Ctrl+C, and then use assert --record.connect --url before any interaction."click the blue Submit button in the contact form", not selectors like "#submit".act command: For example, fill the email and password fields, then click Log In in one prompt. Use separate commands when you need to inspect the intermediate state.assert --prompt "..." for the current page. For a toast, banner, animation, or transition, run record start --output ... in a dedicated terminal, perform the interaction, stop recording with Ctrl+C, wait for the saved-path message, then pass the artifact to assert --record.tap --locate when a reference image is provided: If the user shares a screenshot, icon, or logo and wants that exact visual target, use tap --locate with a multimodal locate JSON object such as { "prompt": "...", "images": [...] } instead of relying only on act --prompt.Example — Dropdown selection:
Example — Form interaction:
Two optional global flags help when Midscene struggles with a task. Put them anywhere in the command (before or after the sub-command); once set, the relevant operations use them by default, so you don't pass a per-call parameter.
--deep-locate — spends an extra round of visual reasoning to pinpoint the target element. Use it when an action interacts with the wrong spot (location drift / offset). It applies to every operation that locates an element, including tap --locate and the locating that happens inside act.--deep-think — plans act with deeper reasoning (richer context and sub-goal decomposition). Use it for complex, multi-step act instructions; it only affects planning.Both trade a little speed for better results, and you can combine them.
In CDP or Bridge mode, keep your usual --cdp / --bridge flags alongside these.
.env file contains MIDSCENE_MODEL_API_KEY=<your-key>.@midscene/* Dependency Version Outdatednpm ls @midscene/web @midscene/core @midscene/shared (or pnpm why @midscene/web).npm view @midscene/web version, npm view @midscene/core version, npm view @midscene/shared version.npm i @midscene/web@latest @midscene/core@latest @midscene/shared@latest.