Drisp Browser (agent-web-interface) is an MCP browser automation server for AI agents, whose core goal is to provide reliable web interaction capabilities with fewer tokens. It drives Chrome based on Puppeteer and the Chrome DevTools Protocol, but what it changes is the page representation that the agent "sees," not the underlying browser execution layer.
[Purpose] Traditional browser tools typically expose the raw DOM, the full accessibility tree, screenshots, or low-level selectors. This is feasible for deterministic scripts, but inefficient for language models, which must spend context and reasoning budget distinguishing useful UI intent from implementation noise. Drisp Browser provides a compact, semantic browser interface for coding agents, browser agents, QA agents, research agents, and automation workflows.
[Main Capabilities] It transforms browser pages into agent-readable snapshots, including: page regions (header, navigation, main content, footer); interactive elements (buttons, links, text boxes, checkboxes, radio buttons, dropdowns); readable content (headings, paragraphs, alerts, labels); element states (visible, enabled, checked, selected, expanded, focused); layout hints (bounding boxes and screen regions); stable element IDs (eid, reusable by subsequent action tools). Agents perform actions through these eids, rather than relying on fragile CSS selectors or repeatedly scanning large DOM trees. The tools cover navigation, interaction, forms, screenshots, canvas inspection, readability, and network activity inspection.
[Usage] Typical agent loop: call navigate with a URL; Drisp Browser returns a compact page snapshot; call find to locate semantic elements (such as a "Log in" button or an email field); use the returned eid to call click, type, select, or press; the server waits for the page to stabilize and returns an updated snapshot; the agent continues based on the changed page state, rather than re-reading the entire DOM.
Most interaction tools return a compact XML state response, including <state> (current page state, title, URL, step number), <meta> (viewport, scroll position, active interaction layer), <baseline> or <diff> (full page state or incremental update), <observations> (transient UI changes such as dialogs, alerts, and prompts), <region> (actionable elements grouped by semantic region), as well as short tags such as <btn>, <inp>, <link>, <chk>, <sel>, <alert>, etc. Subsequent operations should use the id attribute rather than CSS selectors.
For installation, an interactive installer can automatically detect installed AI tools and register the MCP server and agent skills in one step; you can also install for a specific harness, or install only the drisp-browser skill (using npx skills). Manual configuration of Claude Desktop, Cursor, and VS Code is supported; VS Code uses the servers key and requires "type": "stdio". The browser initializes automatically on the first call to a browser tool.
The Chrome launch method can be controlled via DRISP_BROWSER_MODE; DRISP_BROWSER_STEALTH defaults to true, applying only fingerprint patches to the launched Chrome so that it looks like everyday Chrome, but without changing click and input behavior, and is a no-op in user mode. To connect to an existing Chrome profile, enable remote debugging at chrome://inspect/#remote-debugging and allow connections, then launch with DRISP_BROWSER_MODE=user.
[Limitations] It is not a replacement for Puppeteer, not a general-purpose browser, visual testing framework, or scraping framework, nor a CAPTCHA or anti-bot bypass tool; stealth only prevents normal authorized use from being misjudged as a bot. The project is under active development; the public tool interface strives to be simple, but internal APIs and snapshot formats may evolve with actual use. Early internal comparisons showed lower token usage, faster completion, and equal or better success rates compared with Playwright MCP in multi-step navigation tasks, but no formal benchmark suite has been established yet; results depend on the task and should be treated as directional reference only. Running requires Node.js 22.12 or newer, Chrome or Chromium, and an MCP-compatible client (such as Claude Code, Claude Desktop, Cursor, or VS Code).
