AgentWeb

npx add-mcp 'https://mcp.agentweb.us/mcp'

AgentWeb MCP

AgentWeb lets an AI agent complete real tasks on websites — search, order, book, submit, read an account — through typed actions instead of a browser. Every action returns structured output, a verification state and a receipt.

This repository holds the public documentation for the AgentWeb MCP server: how to connect it, what its tools do, example prompts and agent skills. The server itself is hosted; there is nothing to build or run.

Server URL	https://mcp.agentweb.us/mcp
Transport	Streamable HTTP
Auth	OAuth 2.1 (authorization code with PKCE), or an aw_ API key as a bearer token
Docs	https://agentweb.us/docs/quickstart
Pricing	https://agentweb.us/pricing
Status	https://agentweb.us/status

Connect

Claude (claude.ai, desktop, mobile) — Settings → Connectors → Add custom connector, enter https://mcp.agentweb.us/mcp, then sign in when prompted.

ChatGPT — Settings → Connectors → enable Developer Mode → Add custom connector, enter https://mcp.agentweb.us/mcp, then sign in.

Claude Code

claude mcp add --transport http agentweb https://mcp.agentweb.us/mcp

Codex CLI — add to ~/.codex/config.toml:

[mcp_servers.agentweb]
url = "https://mcp.agentweb.us/mcp"

Cursor — add to .cursor/mcp.json:

{ "mcpServers": { "agentweb": { "url": "https://mcp.agentweb.us/mcp" } } }

To have an agent set itself up end to end, give it https://agentweb.us/install.md.

Sign in

There are two ways to authenticate, and both end at the same account.

• OAuth. Clients that support remote MCP sign-in open an AgentWeb sign-in page the first time the connector is used. No key is handled by the user or the agent.
• API key. Clients configured by hand send Authorization: Bearer aw_<key>. An agent can obtain the key inside the conversation: it calls agentweb_start_claim_link_signup, the user opens one link and signs in, and the agent picks the key up by polling. The account-setup skill describes the flow.

Acting on a website as the user (their orders, their inbox) needs a second, separate step: connecting that site account. agentweb_start_connection returns a link the user opens to connect it; the site's credentials are not exposed to the agent.

How a task runs

1. find_actions — describe the task in plain words; get back the supported site, the action, its parameters and a ready call.
2. execute — run that action once and read the returned status, verification and receipt. run does both steps in one call.
3. If the site is not supported yet, request_map asks AgentWeb to learn it and agentweb_map_status reports progress.

read_page fetches any public URL as clean text for context.

What is in this repository

Path	Contents
TOOLS.md	All tools, grouped, with read/write annotations
EXAMPLES.md	Example prompts and what to expect back
skills/	Agent skills for running actions, account setup and requesting a site
server.json	MCP Registry manifest
PRIVACY.md	What the connector sends and stores
SECURITY.md	How to report a vulnerability
SUPPORT.md	Where to get help

License

Copyright 2026 Agent Web Labs, Inc.

The contents of this repository are licensed under the Apache License, Version 2.0. The license does not grant rights to the AgentWeb name or logo. The AgentWeb service is governed by its Terms.