Midscene Automations Skills for Browser

Vision-driven browser automation using Midscene. Operates entirely from screenshots — no DOM or accessibility labels required. Can interact with all visible...

Security Scan

The skill's runtime instructions require external model API keys and execute an unverified npm package via npx, but the registry metadata does not declare those required credentials or a source—this mismatch and the use of npx from an unknown origin are concerning.

Purpose & Capability

The described purpose (vision-driven browser automation using Midscene) aligns with the instructions to run npx @midscene/web and drive a headless Chrome via screenshots. However the skill metadata claims no required environment variables or credentials while the SKILL.md explicitly requires MIDSCENE_MODEL_API_KEY, MIDSCENE_MODEL_NAME, MIDSCENE_MODEL_BASE_URL, and MIDSCENE_MODEL_FAMILY. That metadata/instruction mismatch is incoherent.

Instruction Scope

SKILL.md instructs the agent to run npx CLI commands, take and read screenshots, and rely on a .env file (or system env vars) for model credentials. While reading screenshots is expected, the instructions implicitly expect access to .env and to secrets (API keys) and to execute network-fetched code via npx. The document does not instruct explicit exfiltration, but it gives the agent broad runtime powers (running arbitrary CLI commands from npm, persisting a browser process) which expands its attack surface.

Install Mechanism

There is no install spec in the registry (instruction-only), but the runtime relies on npx @midscene/web@1 which will fetch and run code from the npm registry at runtime. The skill package source and homepage are unknown in registry metadata, increasing risk: running npx pulls arbitrary remote code unless you verify the package/release. This is moderate-to-high risk compared with a pinned, verifiable install source.

Credentials

The SKILL.md requires multiple API-related environment variables (MIDSCENE_MODEL_API_KEY, NAME, BASE_URL, FAMILY, etc.) for external LLM providers, which is reasonable for a vision/LLM-backed tool — but the registry metadata lists no required env vars and no primary credential. The mismatch is problematic: a user would not see declared secrets required before installing. Also the skill suggests storing keys in a local .env file (which the agent may read indirectly via the CLI), so secret handling should be clarified and minimized.

Persistence & Privilege

The skill does not request always:true or other elevated platform privileges. It runs CLI commands that spawn a persistent headless Chrome process across CLI calls (as part of the automation flow), but that is local process behavior, not elevated registry-level privilege.

What to consider before installing

This skill runs an npm CLI (npx @midscene/web@1) and requires model API keys, but the registry metadata doesn't declare those secrets and there's no homepage/source listed. Before installing or running it: (1) ask for the package repository or official homepage and verify the npm package contents (or prefer a published GitHub release), (2) do not supply high-privilege API keys — use limited-scope keys or a quota-limited test project, (3) run the skill first in an isolated/disposable environment, (4) prefer the developer add required env vars to the registry metadata and include integrity/pinned package info, and (5) monitor network and process activity while the skill runs. If you cannot verify the package source, treat it as higher risk and avoid providing real credentials.

Browser Automation

CRITICAL RULES — VIOLATIONS WILL BREAK THE WORKFLOW:

1. 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.
2. 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.
3. 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 act commands may need even longer.
4. 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 @midscene/web@1. Launches a headless Chrome via Puppeteer that persists across CLI calls — no session loss between commands. Each CLI command maps directly to an MCP tool — you (the AI agent) act as the brain, deciding which actions to take based on screenshots.

When to Use

Use this skill when:

• The user wants to browse or navigate to a specific URL
• You need to scrape, extract, or collect data from websites
• You want to verify or test frontend UI behavior
• The user wants screenshots of web pages

If you need to preserve login sessions or work with the user's existing browser tabs, use the Chrome Bridge Automation skill instead.

Prerequisites

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):

MIDSCENE_MODEL_API_KEY="your-api-key"
MIDSCENE_MODEL_NAME="model-name"
MIDSCENE_MODEL_BASE_URL="https://..."
MIDSCENE_MODEL_FAMILY="family-identifier"

Example: Gemini (Gemini-3-Flash)

MIDSCENE_MODEL_API_KEY="your-google-api-key"
MIDSCENE_MODEL_NAME="gemini-3-flash"
MIDSCENE_MODEL_BASE_URL="https://generativelanguage.googleapis.com/v1beta/openai/"
MIDSCENE_MODEL_FAMILY="gemini"

Example: Qwen 3.5

MIDSCENE_MODEL_API_KEY="your-aliyun-api-key"
MIDSCENE_MODEL_NAME="qwen3.5-plus"
MIDSCENE_MODEL_BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
MIDSCENE_MODEL_FAMILY="qwen3.5"
MIDSCENE_MODEL_REASONING_ENABLED="false"
# If using OpenRouter, set:
# MIDSCENE_MODEL_API_KEY="your-openrouter-api-key"
# MIDSCENE_MODEL_NAME="qwen/qwen3.5-plus"
# MIDSCENE_MODEL_BASE_URL="https://openrouter.ai/api/v1"

Example: Doubao Seed 2.0 Lite

MIDSCENE_MODEL_API_KEY="your-doubao-api-key"
MIDSCENE_MODEL_NAME="doubao-seed-2-0-lite"
MIDSCENE_MODEL_BASE_URL="https://ark.cn-beijing.volces.com/api/v3"
MIDSCENE_MODEL_FAMILY="doubao-seed"

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.

Commands

Connect to a Web Page

npx @midscene/web@1 connect --url https://example.com

Take Screenshot

npx @midscene/web@1 take_screenshot

After taking a screenshot, read the saved image file to understand the current page state before deciding the next action.

Perform 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:

# specific instructions
npx @midscene/web@1 act --prompt "click the Login button and fill in the email field with '[e-mail]'"
npx @midscene/web@1 act --prompt "scroll down and click the Submit button"

# or target-driven instructions
npx @midscene/web@1 act --prompt "click the country dropdown and select Japan"

Disconnect

Disconnect from the page but keep the browser running:

npx @midscene/web@1 disconnect

Close Browser

Close the browser completely when finished:

npx @midscene/web@1 close

Workflow Pattern

The browser persists across CLI calls via a background Chrome process. Follow this pattern:

1. Connect to a URL to open a new tab
2. Take screenshot to see the current state, make sure the page is loaded.
3. Execute action using act to perform the desired action or target-driven instructions.
4. Close the browser when done (or disconnect to keep it for later)
5. Report results — summarize what was accomplished, present key findings and data extracted during the task, and list any generated files (screenshots, logs, etc.) with their paths

Best Practices

1. Always connect first: Navigate to the target URL with connect --url before any interaction.
2. Be specific about UI elements: Instead of "the button", say "the blue Submit button in the contact form".
3. Use natural language: Describe what you see on the page, not CSS selectors. Say "the red Buy Now button" instead of "#buy-btn".
4. Handle loading states: After navigation or actions that trigger page loads, take a screenshot to verify the page has loaded.
5. Close when done: Use close to shut down the browser and free resources.
6. Never run in background: Every midscene command must run synchronously — background execution breaks the screenshot-analyze-act loop.
7. Batch related operations into a single act command: When performing consecutive operations within the same page, combine them into one act prompt instead of splitting them into separate commands. For example, "fill in the email and password fields, then click the Login button" should be a single act call, not three. This reduces round-trips, avoids unnecessary screenshot-analyze cycles, and is significantly faster.
8. Always report results after completion: After finishing the automation task, you MUST proactively present the results to the user without waiting for them to ask. This includes: (1) the answer to the user's original question or the outcome of the requested task, (2) key data extracted or observed during execution, (3) screenshots and other generated files with their paths, (4) a brief summary of steps taken. Do NOT silently finish after the last automation command — the user expects complete results in a single interaction.

Example — Dropdown selection:

npx @midscene/web@1 act --prompt "click the country dropdown and select Japan"
npx @midscene/web@1 take_screenshot

Example — Form interaction:

npx @midscene/web@1 act --prompt "fill in the email field with '[e-mail]' and the password field with 'pass123', then click the Log In button"
npx @midscene/web@1 take_screenshot

Troubleshooting

Connection Failures

• Ensure Chrome/Chromium is installed on the system (Puppeteer downloads its own by default).
• Check that no firewall blocks local Chrome debugging ports.

API Key Errors

• Check .env file contains MIDSCENE_MODEL_API_KEY=<your-key>.
• Verify the key is valid for the configured model provider.

Timeouts

• Web pages may take time to load. After connecting, take a screenshot to verify readiness before interacting.
• For slow pages, wait briefly between steps.

Screenshots Not Displaying

• The screenshot path is an absolute path to a local file. Use the Read tool to view it.