Rindler
OverviewQuick StartCLICore toolsTool referenceError codesSign in to gated sitesMapping new sites

Agent quickstart

Start a structured website session

No API key, no custom functions. Connect once, then tell your agent to do a task on any supported site.

Install the MCP server

One command, run in your project root:

curl https://rindler.ai/install | sh

This runs a short installer that points your runtime at https://mcp.rindler.ai. It writes .mcp.json (or merges Rindler into an existing one), backs up the previous file, is safe to re-run, and never touches a config it cannot parse. It covers Claude Code, Cursor, Devin Desktop, Claude Desktop, and Codex. In Codex you will be asked to approve running the installer, since it executes a script rather than only writing a file.

To wire a runtime by hand, add the server to its MCP config:

{
  "mcpServers": {
    "rindler": {
      "type": "http",
      "url": "https://mcp.rindler.ai"
    }
  }
}

Connect and authenticate

Restart your agent so it loads the new tools. The first tool call returns an authorization link your runtime opens in the browser, where you approve access over OAuth 2.0 PKCE. No API key, nothing to paste.

Tokens refresh automatically. If a runtime ever needs to re-authenticate, it reruns the same browser login.

Run your first task

With Rindler connected, just tell your agent what to do on a supported site. For example: "Tell me the price of eggs on Instacart in my area." The agent opens a session, runs the right actions, reads the result back as structured data, and closes the session. You do not write any glue code.

The tools it uses (start_session, dispatch_action, extract_content, close_session) are documented in Rindler MCP.

When a site isn't supported

Rindler only drives sites in its catalog. For an unmapped site, start_session returns screen: "unsupported_site" with supported: false. That is a signal, not an error: the agent can fall back to native browsing instead of improvising against raw HTML.