agentbrowse

Key Features

• •Read page as token-bounded markdown: The 'read' command returns the current page content as clean markdown, respecting token limits (--max-chars) for LLM consumption. This allows AI agents to efficiently ingest web text without HTML noise, directly usable for context generation.
• •Accessibility-tree snapshot with stable refs: The 'snapshot' command outputs every actionable element with a stable [ref], role, name, and state. Agents can identify interactive elements robustly, and the 'click' command uses these refs to perform reliable clicks without fragile CSS selectors.
• •Click by visible text, ref, or number: The 'click' command supports multiple targeting methods: visible text, a number from links/find output, or a snapshot ref. This gives agents flexibility to interact with buttons, links, or any clickable element using the most contextually appropriate method.
• •Persistent browser session across commands: A background Playwright daemon maintains one live page per session. Commands like open, read, click all hit the same daemon, so state (cookies, navigation) persists. Sessions are isolated by ID, allowing separate contexts.
• •Login command for SSO and MFA: The 'login' command opens a real browser window for a human to authenticate (including SSO, MFA, captchas). The session is saved locally and reused headlessly. Credentials are never exposed in CLI arguments, solving a key security and usability challenge.
• •Site manifests for high-level commands: Optional site.agent.json files define named commands with ordered steps, interpolation of pages/selectors/args/env. Agents can run 'agentbrowse --site ./notion.agent.json search "query"' to execute domain-specific actions, reducing complexity.
• •Skill installation for Claude Code and others: Running 'npx agentbrowse skill' auto-detects agents like Claude Code, Codex, Cursor, Gemini, Windsurf and writes native format config files (SKILL.md, .cursor/rules, AGENTS.md). This makes agentbrowse available automatically for web tasks without prompting.

Use Cases

• •Research and documentation gathering: An AI agent like Claude Code is assigned to research a topic. It runs 'agentbrowse open https://docs.example.com', then 'agentbrowse read --max-chars 8000' to extract relevant paragraphs as markdown, which it uses to answer questions in a coding session. Outcome: the agent can incorporate live web content into its reasoning without leaving the terminal.
• •Automated form submission for testing: A QA engineer scripts a CI pipeline that tests a web app's registration flow. Commands like 'agentbrowse fill -f name=test password=123' and 'agentbrowse submit' are executed sequentially. The persistent session ensures the same form and cookies are used. Outcome: repeatable, no-GUI testing of web forms.
• •Scraping paginated search results: An agent needs to collect data from multiple pages of a search results site. After 'agentbrowse open https://search.com?q=ai', it runs 'agentbrowse links' to find page numbers, then 'agentbrowse click 3' to go to page 3, and continues. The markdown output is saved or piped. Outcome: systematic data extraction across pages with minimal commands.
• •Authenticated dashboard monitoring: A DevOps engineer uses agentbrowse to log into a cloud dashboard via SSO. They run 'agentbrowse login https://dashboard.company.com', complete the authentication in a real browser, then later 'agentbrowse read' to extract status metrics. Outcome: headless monitoring of authenticated dashboards without storing credentials in scripts.
• •Site-specific command execution with manifests: A developer working with Notion creates a site.agent.json manifest with a 'search' command. They run 'agentbrowse --site ./notion.agent.json search "Q4 roadmap"', which executes a sequence of opens, clicks, and reads to retrieve the result. Outcome: domain-specific automation with natural language-like commands.

Who is this for?

Developers using AI coding agents such as Claude Code, Codex, Cursor, Gemini, and Windsurf who need to automate web interactions from the terminal. Also suitable for QA engineers and DevOps professionals who require headless browser automation for testing, monitoring, and data extraction. Anyone building workflows where AI agents must read, navigate, fill forms, or authenticate on web pages will find agentbrowse essential. It targets technical users comfortable with npm, Playwright, and CLI tools who want to eliminate manual clicking and enable their agents to treat the web as a command-line resource.