Skip to main content
Pick how your agent connects to Agent Vault in under a minute, then go to its setup page. If you are unsure, start with the MCP server. The dashboard’s Overview page generates the finished setup. Where your deployment runs a hosted MCP server, OpenClaw, Claude Code, Hermes and Codex each have a Sign in route, a line to paste with no API key, and a Use an API key route, the self-run block with your key filled in. Meta Muse, Grok Bot, Claude Desktop and Antigravity only sign in, and Claude Desktop is added in its settings rather than with a pasted line. For the agents with their own setup page, see Connect your agent.

​
MCP server

@agentvault/mcp gives your agent 29 tools, such as get_token_catalog, fill_checkout and request_challenge_handoff. It drives one page at a time, reused across calls, which suits conversational agents and one-off errands. You connect it in one of two ways:
  • Hosted, where your deployment runs one: your host adds a remote server and you sign in with your Agent Vault account. There is no API key and nothing to install.
  • Self-run: your host launches the server as a local stdio process, with your API key in its environment.
See MCP setup and the tool reference.

​
Playwright SDK

@agentvault/playwright is a Playwright-shaped browser object model that talks to the proxy over a WebSocket. You get full contexts, pages, popups and frame locators, plus your own retry logic, which suits scripted runs and batch jobs. It runs on Bun, or Node 20 or later. See Install the SDK.

​
Agent skill

The skill is a file, not a client. It teaches an agent to use only mock tokens, never invent a password, fill card fields in the right order, and hand challenges to you. A companion checkout recipes skill teaches it to look up a store in the registry first.

​
What every option shares

The same protections. The rules belong to the token, so a fill through fill_checkout is checked exactly like one through page.fill(), whichever way your agent connects.

​
What you need

The SDK and the self-run MCP server need: The hosted MCP server needs none of these. It needs:
  • An MCP client that supports remote servers over streamable HTTP with OAuth. You sign in with your Agent Vault account, and the client keeps the token.
  • Access chosen on Connected apps. A new connection can use nothing until you click Choose access on its row on Connected apps and then Save access. See Connect to the hosted server.
Partly. An agent in any language can read its token catalog and the recipe registry over HTTP. Filling fields goes through the proxy’s WebSocket protocol, and the MCP server and the SDK are its two supported clients. See HTTP API.
Next: MCP setup to add Agent Vault to your MCP host.