Skip to content
Last updated • May 2026

Customaise MCP. Your AI agent, in your real browser.

Customaise MCP is a local bridge between any AI coding agent and your real, signed-in Chrome session. Your agent's "brain" stays in your IDE; its "hands" run in the browser where you are already logged in, joined by a WebSocket on your own machine. Once connected, the agent gets 18 built-in tools plus every WebMCP tool your AgentScripts register, and every sensitive call passes a consent gate the agent cannot forge.

Key takeaways

  • Customaise MCP is a local bridge: your IDE agent on one side, your real signed-in Chrome on the other, joined by a WebSocket on localhost:4050.
  • Add it to Cursor, Claude Code, Codex, Antigravity, or Kiro with one config line, or install the one-click .mcpb bundle into Claude Desktop.
  • Bring Your Own Session: tools run in the browser you are already logged into, so no API keys, no OAuth dance, no headless login, no scraping.
  • Your agent gets 18 built-in tools (read DOM, screenshot any tab, manage scripts, control tabs) plus every tool your installed AgentScripts expose.
  • Free up to 50 tool calls a day and 150 a rolling week, unlimited on Power User. MCP calls never spend your in-browser AI token allowance.

Two browsers, one session.

Customaise is split across two browsers on purpose. The brain reasons in your IDE; the hands act where you are logged in.

The brain

Your IDE agent

Cursor, Claude Code, Codex, and other MCP clients live in your IDE. The agent reasons and decides here. Its built-in browser is not where Customaise runs.

The bridge

localhost:4050

A local server, spawned by your IDE's MCP client with one npx command, speaks the v2 bridge protocol to your extension over a WebSocket on your own machine.

The hands

Your signed-in Chrome

The extension runs each tool in the tab you are already authenticated in. The agent never sees your cookies or tokens; it only calls tools that act inside your session.

The unlock

Bring Your Own Session

A headless browser logs in fresh every run or juggles stored cookies. A REST integration needs an API key per site, if the site even ships one. Customaise sidesteps both: the agent's tools execute inside the Chrome you already use, with the sessions you have already established. The bank, the CRM, the internal Jira, the dashboard with no public API at all, each becomes addressable the moment you open it and sign in normally.

The agent operates as you, with your permission, and never holds your credentials. It calls a named tool; the tool runs in your authenticated tab and returns structured JSON. That is the difference between an agent that can read a screenshot of your inbox and one that can actually act in it.

Two minutes

Connect your agent

Install the Customaise extension and sign in from the popup. Then point your IDE agent at the server. The same npx command works for Cursor, Claude Code, Codex, Antigravity, and Kiro:

{
  "mcpServers": {
    "customaise": {
      "command": "npx",
      "args": ["-y", "@customaise/mcp"]
    }
  }
}

Cursor reads .cursor/mcp.json, Claude Desktop reads claude_desktop_config.json, Codex uses TOML (command = "npx"). Same server, same arguments. Keep Customaise running in your daily browser, not the IDE's built-in one, so tools land in your signed-in session.

Claude Desktop: one-click bundle

Prefer not to edit a config file? Download the .mcpb bundle and open it in Claude Desktop. It registers the same server in one click. The version matches the published @customaise/mcp package.

Download customaise.mcpb
18 built-in tools

What your agent can do

The bridge ships eighteen built-in tools for the browser itself, and adds every WebMCP tool your installed AgentScripts register on top.

Script lifecycle
list_scriptsimport_scriptexport_scriptdelete_scripttoggle_script
Browser context
get_page_contextget_console_contextlist_tabs
Tab control
open_tabclose_tabfocus_tabreload_tab
Visual DOM targeting
get_selected_elementstake_screenshot
WebMCP agent tools
list_webmcp_toolscall_webmcp_tool
UI control and batch
toggle_uisync_scripts
How Customaise teaches your agent

Conventions, read before building

Alongside the tools, the server exposes five MCP resources. Two are full handbooks that define exactly how Customaise expects scripts to be written, down to working recipes. A well-behaved agent reads the relevant handbook before it touches a script, which is why an agent connected to Customaise already knows how to build a UserScript, an AgentScript, or a CM_devtools script correctly, with no instructions pasted in.

customaise://scripts
A live JSON list of every script the extension manages, with IDs, names, enabled state, and match patterns.
customaise://scripts/{id}
The full source and metadata for one script, so the agent can read before it edits.
customaise://conventions
A router that points the agent at the right handbook for the script type it is working on.
customaise://userscript-conventions
The full UserScript reference: file structure, the IIFE pattern, the GM_ APIs, symbol-level editing, @match and @namespace rules.
customaise://agentscript-conventions
The full AgentScript reference: the metadata block, @webmcp permission declarations, navigator.modelContext, the consent model, CM_devtools, and a set of working recipes.

The page becomes a REPL

The deeper habit the conventions teach: the agent does not have to write one big script that works first try. It ships a small read-only tool that surfaces page state, calls it, reads the structured JSON, edits the script, re-exports, and calls again, a tight loop against the live page, then throws the diagnostic tool away when the real one works. The agent debugs itself. That loop is why scripts for messy, undocumented sites come together in minutes.

What you get

Free to taste, unlimited on Power User

Free, signed in

  • Connect any MCP agent to your browser
  • All 18 built-in tools plus your AgentScript tools
  • 50 tool calls per day, 150 per rolling week
  • In-browser AI chat can call the same tools

Power User

  • Unlimited MCP tool calls
  • Cloud script publishing and sharing
  • Remote HITL approvals on your phone
  • A monthly in-browser AI token allowance for AI-built scripts

MCP tool calls and the in-browser AI token allowance are separate meters; calling tools never spends your AI tokens. See the pricing page for current rates.

Frequently asked

Any MCP-capable client. Out of the box: Cursor, Claude Code, OpenAI Codex, Google Antigravity, Amazon Kiro, and Windsurf connect through their MCP config, and Claude Desktop installs the one-click .mcpb bundle. Each uses the same npx command under the hood, so a new MCP host usually works by pointing it at "npx -y @customaise/mcp".

No, and that is the whole point. Tools run inside your real Chrome session, so they use the cookies and logins you already have. You sign in to Customaise once from the extension popup; there are no per-site API keys, no service accounts, and no headless browser to authenticate. Customaise calls this Bring Your Own Session: any web app you are logged into becomes addressable by your agent.

Every WebMCP tool an AgentScript registers carries a permission: allow runs immediately, prompt shows an in-browser consent modal with the tool name and arguments before every call, and deny blocks it (undeclared tools default to deny). The gate is enforced in the extension service worker on every call path, the IDE client, the in-extension AI, or any in-page agent, so a tool marked prompt always shows the modal regardless of who calls it. The agent cannot forge or click that modal. Power User can mirror it to a phone.

Because the value is your authenticated session. Customaise is deliberately split across two browsers: the agent reasons in your IDE, but its tools must execute where you are logged in, your daily Chrome, Edge, or Brave. Running Customaise in the IDE's throwaway browser would lose the very sessions that make the tools useful.

No. The in-browser AI token allowance is for the AI editor that writes scripts from a sentence. MCP tool calls are a separate meter: free up to a daily and weekly cap, unlimited on Power User. The two never draw from each other.

Nothing different at runtime; it is the same server. The .mcpb is Claude Desktop's one-click install format: download it, open it in Claude Desktop, and the server is registered without editing a JSON file by hand. IDE agents like Cursor and Claude Code use the config line instead.

The current MCP server speaks the v2 bridge protocol, which needs the Customaise extension at version 1.2.3 or newer (any current Web Store build is fine). Older extensions still connect with reduced behaviour. You also need to be signed in to Customaise in the extension.

AgentScripts

User-built WebMCP tools your AI agent calls on the sites you already use.

Read more

UserScripts

Reshape any site with a small script, written from a sentence.

Read more

Chrome DevTools Access

Give a script the Chrome DevTools Protocol when the DOM is not enough.

Read more

See Customaise in action

How it all fits together: AI editing, the MCP bridge, cross-device sync, and the built-in browser tools.

See how it works

Point your agent at your browser.

Add Customaise to Chrome, sign in, and drop the config line into your IDE. Your agent gains the whole browser as a toolset.

Add to ChromeBack to Knowledge Base