Agent Web Interface

What lespaceman/agent-web-interface MCP server does

The lespaceman/agent-web-interface MCP server gives an AI agent a semantic interface to Chrome. Rather than returning a complete DOM or relying on screenshots as the primary representation, it describes visible page regions, readable text, interactive controls, form context, element state, layout hints, and stable element IDs. The result is intended to reduce the amount of browser detail an agent must inspect before choosing an action.

The server supports browser workflows such as opening a URL, locating a meaningful control, interacting with a form, and checking the resulting page state. It is aimed at coding, browser, QA, research, and automation agents that need to operate existing web interfaces. It is not presented as a replacement for Puppeteer, a general-purpose browser, a visual testing framework, a scraping framework, or a CAPTCHA bypass tool.

How it works

The lespaceman/agent-web-interface MCP server receives browser tool calls over MCP and controls Chrome through Puppeteer and CDP. It converts the current page into a compact semantic snapshot. Elements such as buttons, links, textboxes, checkboxes, radio buttons, and comboboxes can receive stable IDs, called eid values, which the agent can pass to later action tools.

A typical interaction starts with navigation. The response provides the page title, URL, regions, readable content, controls, and relevant state such as whether an element is enabled, selected, expanded, checked, or required. The agent can then find an element and use an ID with an action such as clicking, typing, selecting, or pressing a key. After the action, the server waits for the page to stabilize and returns either a new snapshot or a focused change.

Navigation establishes a new baseline, so IDs from the previous page should not automatically be reused. For mutations, the response can describe only the changed area. This observation-action relationship lets an agent continue from the latest semantic state instead of rescanning a large page representation after every step.

Setup and configuration

The project publishes an npm package named @drisp/browser-mcp. A local installation can therefore use the package runner command npx -y @drisp/browser-mcp, subject to the project’s runtime requirements. The provided material identifies Chrome or Chromium as the browser being controlled and Puppeteer plus Chrome DevTools Protocol as the execution layer.

The supplied documentation does not specify environment variables, authentication settings, operating-system requirements, or a detailed Chrome launch configuration. Websites that require their own login or session may still require the agent’s browser session to be prepared separately; no server-level credential mechanism is described here.

Tools and capabilities

Documented capabilities include:

• Navigate to web pages and receive semantic page snapshots.
• Find controls or content using semantic information.
• Click, type, select, and press keys using stable element IDs.
• Inspect visible regions, readable content, forms, and interactive elements.
• Track element state such as visibility, enabled status, focus, selection, expansion, and checked state.
• Receive screenshots, canvas inspection, and network activity information.
• Observe updated state or focused diffs after browser actions.

The snapshots may also expose bounding boxes and screen zones as layout hints. Stable IDs connect the observation returned by the server to the next action, while transient observations can report events such as a status message appearing after a save operation.

Limitations and notes

The lespaceman/agent-web-interface MCP server changes the agent-facing representation of browser state; it does not replace Chrome, Puppeteer, or CDP. Its documented scope is browser interaction through semantic observations and actions, not general scraping or visual regression testing. The material also explicitly excludes CAPTCHA and anti-bot bypass functionality.

Because element IDs belong to the current semantic page state, a navigation or substantial page change can invalidate earlier IDs. Agents should use the latest returned snapshot or diff before acting. The provided material does not establish support for a particular MCP client, hosted endpoint, browser vendor other than Chrome or Chromium, or authentication provider.