Works with the AI Agents you already use
Browser Agent MCP Server
Let your agent drive real web pages via accessibility snapshots and stable element handles — fewer tokens and no fragile coordinate guessing.
Free
Browser Agent MCP Server
Example session with this skill installed
I'm building a support bot in Cursor that needs to log into our internal help desk and close resolved tickets. Screenshot-based attempts keep clicking the wrong buttons and burn tokens. Set up the accessibility-driven browser MCP server for me: register it in my mcp.json, then give me the exact navigate-snapshot-act loop my agent should follow. Context: standard React app with labeled inputs, the agent runs headless on Ubuntu behind an HTTP proxy, and we need to keep login state between sessions. Include the launch flags and any environment variables involved.
- Read your context and instructions
- Compiled the browser agent mcp
Summary
Your agent now drives the help desk through accessibility snapshots and handle-based clicks — no coordinate guessing. The server runs headless behind your proxy, persists login state in a dedicated profile, and registers as a single MCP entry in Cursor.
{
"mcpServers": {
"web-agent": {
"command": "npx",
"args": ["@agent-infra/mcp-server-browser@latest", "--headless", "--proxy-server", "http://proxy.internal:3128", "--user-data-dir", "./.agent-profile", "--output-dir", "./.agent-out"]
}
}
}
| Step | Tool | Rule |
|---|---|---|
| 1 | browser_navigate | Open the ticket URL once |
| 2 | browser_get_clickable_elements | One snapshot per page state |
| 3 | click / fill / select | Act by handle index, never x/y |
| 4 | re-snapshot | Only after the page actually changed |
| 5 | screenshot | Fallback when labels are missing |
Next steps
- Dry-run one resolved ticket before letting the bot close real ones
- Add
--visiononly for canvas widgets; return to structured tools if a vision click misses - Keep
.agent-profileout of git and inspect the download dir before sharing files
Connects securely to your tools. The creator never sees your data.
About this skill
Give your language model a real browser — without the pixel guessing. Instead of streaming screenshots to a vision model and hoping coordinates land, this reads pages through Puppeteer's accessibility tree and hands the model stable element handles. Clicks, form fills, scrolls and JS evaluation become repeatable at a fraction of the token cost. What you get: - Three connection modes: local stdio, a shared HTTP/SSE endpoint for workers, or in-process embedding for TypeScript runtimes - Full coverage: navigation, tabs, text and markdown extraction, interaction, screenshots, console JS, plus a vision fallback for canvas apps - Per-request headers for viewport, user agent and coordinate scaling — no restarts - A runbook for Docker displays, proxies, profile security and common failures Ship deterministic, cheaper, easier-to-debug browser agents.
Canvas-based or label-poor pages fall back to slower vision clicks, which can miss when coordinate conventions differ; the HTTP/SSE endpoint ships unauthenticated, so public exposure requires your own reverse-proxy auth.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
This skill has been used 7 times.
Trust & safety
Security scan
Security scan passed 2 months ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts
File Scopes
Needs shell and network to install and launch the browser MCP server and drive real pages, plus write access for client config files and browser profiles.