> ## Documentation Index > Fetch the complete documentation index at: https://docs.apostate.dev/llms.txt > Use this file to discover all available pages before exploring further. # AI agents > Give Claude Code, Codex or your own agent a browser that runs headless and reads as a normal desktop machine. An agent that browses needs a browser the web treats as an ordinary visitor. Headless Chromium driven by stock Playwright gives itself away: in the [test suite](/testing/results) it failed 12 rows on bot.sannysoft.com, and FingerprintJS Pro flagged it as a bot. Apostate runs headless and presents a Windows, macOS or Linux machine whose GPU, screen, fonts, voices, locale and timezone agree with each other. For an agent that means: * **No window.** The browser runs headless on a server or on your laptop, so an agent can browse while you work, and nothing pops up. * **Checks a person passes.** When a site shows a Cloudflare Turnstile checkbox, the agent clicks it once, as a person would. On public Turnstile demo pages, headless Apostate received a token where headless stock Chromium did not. [Test results](/testing/results) has the dated measurements. * **Sessions that last.** A persistent profile keeps cookies, logins and the same machine between sessions, so a site sees a returning visitor instead of a new device every run. ## Two ways to connect | | MCP server | Scripts | | ------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | | What the agent does | Calls browser tools one step at a time: navigate, snapshot, click, type | Writes a Python or Node script with the Apostate package and runs it | | Good for | Interactive tasks, exploring a site, filling a form once | Repeatable jobs, many pages, anything you want to keep as code | | Setup | One line in the agent's MCP config | `pip install apostate` and a skill that shows the agent how | | Pages | [MCP server](/agents/mcp) | [Claude Code](/agents/claude-code#scripts-and-the-apostate-skill), [Codex](/agents/codex#scripts-and-skills) | Most agents benefit from both. Use the MCP server for step-by-step work, and the skill so the agent can write a script when a task repeats. ## Pick your agent `claude mcp add`, permissions, the Apostate skill, and headless runs with `claude -p`. `config.toml`, tool approval for `codex exec`, environment variables and skills. MCP configuration for Cursor, Gemini CLI, VS Code, OpenCode and Windsurf. A browsing agent on the Claude API with Apostate's browser tools. browser-use connected to Apostate over CDP. Agents on a Linux server or in a container, with no display. ## Sandboxes Claude Code and Codex can run shell commands in a sandbox. Chromium does not start inside either one: on macOS it fails with `bootstrap_check_in ... MachPortRendezvousServer ... Permission denied`, and under Codex on Linux it stops with `SIGTRAP`. MCP servers are started by the agent itself, outside its sandbox, so a browser launched by the Apostate MCP server works whatever the sandbox setting. For scripts, allow the command to run outside the sandbox, or turn the sandbox off for that project. ## Safety An agent that reads web pages reads text written by strangers, and some of it is written to steer agents. * Give the agent its own profile. Do not point it at a profile that holds your personal logins unless the task needs them. * Keep tool approval on for actions that spend money, post, send or delete. * Tell the agent that page content is data, not instructions. The [Apostate skill](/agents/claude-code#scripts-and-the-apostate-skill) and the [Claude API example](/agents/claude-api) do. * Follow each site's terms. Apostate changes what a site can tell about the machine; it does not change what you are allowed to do there.