agentbrowse

Key Features

• •Token-bounded Markdown Reading: The read command outputs the current page as token-bounded markdown, with options --max-chars and --page to control content length. This ensures AI agents receive text within their token limits, making it easy to parse and process without exceeding context windows.
• •Link Navigation with Filters: The links command displays numbered, followable links on a page, supporting an optional --filter to narrow results. Agents can then click a link by its number, enabling efficient navigation to relevant sections without scanning full HTML.
• •Accessibility Tree Snapshot: The snapshot command provides an accessibility-tree view of the page, showing every actionable element with a stable reference, role, name, and state. This gives agents a reliable way to identify and interact with page elements using refs.
• •Find and Click by Text or Ref: The find command locates elements by visible text, falling back to accessible names, and returns numbers that can be used with click. The click command accepts refs, text, numbers, or CSS selectors, offering multiple ways to trigger actions.
• •Form Filling and Submission: The type and fill commands allow entering text into fields identified by CSS selector or name, with fill accepting multiple key-value pairs. The submit command triggers form submission, enabling agents to complete forms programmatically.
• •Real Browser Login for Authentication: The login command opens a real browser window for the human to authenticate on sites with SSO, MFA, or captchas. The session is saved locally, and the agent can operate headlessly with the authenticated state, avoiding credential exposure.
• •Site Manifests for Custom Workflows: Site manifests (site.agent.json) define high-level named commands for specific sites, using pages, selectors, and ordered steps with variable interpolation. This allows agents to execute complex workflows like searching or data extraction with a single command.
• •One-Command Skill Installation: The agentbrowse skill command auto-detects coding agents (Claude Code, Codex, Cursor, Gemini, Windsurf) in the project and writes native configuration files. This makes the tool available to the agent without manual prompting, streamlining integration.

Use Cases

• •An AI coding agent researching a technical documentation site: it opens the page, reads the content as markdown, follows links to subpages, and extracts code examples. Outcome: the agent autonomously gathers information without leaving the command line.
• •An agent tasked with filling a web form to submit a support ticket: it uses type and fill commands to enter details, clicks submit, and reads the confirmation page. Outcome: repetitive form submissions are automated within the agent's workflow.
• •A developer using multiple authenticated sessions: agentbrowse login handles SSO/MFA for each account, and --session isolates cookies. Outcome: the agent can switch between personal and work contexts without manual re-login.
• •An agent monitoring a dashboard: it opens the page, reads current metrics as JSON with --json, and processes the data. Outcome: real-time data extraction becomes part of automated reporting scripts.
• •A QA engineer integrating web testing into agent workflows: the snapshot command provides stable element references for clicking and verifying state. Outcome: reliable element interaction without fragile CSS selectors.
• •An AI assistant helping with online purchases: it navigates product pages, reads descriptions, fills cart, and checks out, all via CLI commands. Outcome: the agent performs multi-step e-commerce tasks autonomously.

Who is this for?

Developers and AI practitioners using coding agents such as Claude Code, Codex, Cursor, Gemini, and Windsurf. Specifically, software engineers who build automated workflows that require web data extraction, form submission, authentication handling, and multi-step browsing from the terminal. Also beneficial for QA automation engineers and web scrapers who want to integrate web interactions into agent-driven pipelines. The tool is designed for anyone who wants to extend their AI coding agent's capabilities to interact with any website without manual browser manipulation.