ie-mode-mcp

MCP Server for operating legacy web applications running in Microsoft Edge's IE mode from AI agents via MCP (Model Context Protocol).

AI Agent ──(MCP / stdio)──> ie-mode-mcp ──> BrowserManager ──> selenium-webdriver
                                                                     │
                                                          IEDriverServer.exe
                                                                     │
                                                     Microsoft Edge (IE Mode)
                                                                     │
                                                       Legacy Web Application

• Configured only with Node.js 22 / TypeScript / selenium-webdriver (no HTTP Server, DB, DI, Logging Framework)

• MCP Transport is stdio only

• Browser session is only one, WebDriver operations are fully sequential

• Does not return full HTML, inspect_page returns summarized screen information for LLM

• No approval flow. Operations are executed when the Tool is called.

Table of Contents

1. Quick Start

2. Prerequisites

3. Windows Pre-configuration

4. Installation and Build

5. Environment Variables

6. Startup Methods

7. Registration with AI Agent

8. Tool Reference

9. Usage Examples

10. Errors and Troubleshooting

11. Logging

12. Troubleshooting

Related MCP server: Selenium MCP Server

1. Quick Start

Run the following on Windows.

git clone https://github.com/sumikof/iedriver-mcp.git
cd iedriver-mcp
npm install
npm run build

# IEDriverServer.exe のパスと、遷移を許可する Origin を指定して起動
$env:IE_MCP_DRIVER_PATH = "C:\tools\IEDriverServer.exe"
$env:IE_MCP_ALLOWED_ORIGINS = "http://legacy01.local"
node dist/index.js

If {"level":"info","event":"started","transport":"stdio"} is output to stderr, startup is successful. Normally, do not start manually; let it auto-start from the AI Agent's MCP settings.

2. Prerequisites

• Download IEDriverServer.exe from the Selenium download page and place it in any folder (e.g., C:\tools\). Since the 64-bit version has known limitations, Selenium officially recommends using the 32-bit version.

• IEDriver is affected by GUI, window focus, and native events, so use in a dedicated Windows VM or dedicated Windows session is recommended.

• Configuration to run the browser on a Windows Service (Session 0) is not assumed.

• MCP Server and IEDriver / Edge run in the same Windows environment.

3. Windows Pre-configuration

IEDriver is strongly affected by environment settings. Complete the manual configuration first before starting the MCP Server.

3.1 Enable Edge IE Mode

First, manually verify with Edge that the target site can be opened in IE mode. IE mode is enabled via one of the following policies (under Software\Policies\Microsoft\Edge).

Specific configuration depends on organizational policy, so check with the Microsoft IE mode documentation and your organization's administrator. Apply the latest updates to Windows / Edge.

3.2 IEDriver Required Settings

If the Protected Mode settings are not unified, browser_start will fail. IEDriver's introduceFlakinessByIgnoringProtectedModeSettings is not used because it makes behavior unstable.

4. Installation and Build

npm install     # 依存パッケージの取得
npm run build   # TypeScript を dist/ へビルド

The output is dist/index.js. After building, it can also be started with npm start (= node dist/index.js).

5. Environment Variables

Configuration files (YAML/JSON) are not used; settings are configured only via environment variables.

IE_MCP_EDGE_PATH=C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe
IE_MCP_DRIVER_PATH=C:\tools\IEDriverServer.exe
IE_MCP_ALLOWED_ORIGINS=http://legacy01.local,http://legacy02.local
IE_MCP_TIMEOUT_MS=10000

• IE Driver 4.5.0 and later automatically detects Edge in environments without IE (Windows 11 default), so IE_MCP_EDGE_PATH is usually unnecessary. Only specify explicitly when auto-detection fails.

• To prioritize operational reproducibility, it is recommended to explicitly specify IE_MCP_DRIVER_PATH.

• IE_MCP_ALLOWED_ORIGINS is a simple restriction to prevent accidental operations, and is determined by exact match of Origin (scheme + host + port). Path-based restrictions are not performed.

6. Startup Methods

Manual Startup (for verification)

PowerShell:

$env:IE_MCP_DRIVER_PATH = "C:\tools\IEDriverServer.exe"
$env:IE_MCP_ALLOWED_ORIGINS = "http://legacy01.local"
node dist/index.js

Command Prompt:

set IE_MCP_DRIVER_PATH=C:\tools\IEDriverServer.exe
set IE_MCP_ALLOWED_ORIGINS=http://legacy01.local
node dist\index.js

Listens for connections from the client via stdio. Since standard input/output is used for the MCP protocol, there is no response to keyboard input in this state (normal). All logs are output to stderr. Exit with Ctrl+C (browser also closes automatically).

Note: Starting the MCP Server alone does not start the browser. The browser starts when the Agent calls browser_start.

Normal Operation

The AI Agent (MCP client) starts this Server as a child process. Manual startup is not required. Configure as described in the next chapter.

7. Registration with AI Agent

Add the following to the MCP client's configuration file.

{
  "mcpServers": {
    "ie-mode": {
      "command": "node",
      "args": ["C:\\ie-mode-mcp\\dist\\index.js"],
      "env": {
        "IE_MCP_DRIVER_PATH": "C:\\tools\\IEDriverServer.exe",
        "IE_MCP_EDGE_PATH": "C:\\Program Files (x86)\\Microsoft\\Edge\\Application\\msedge.exe",
        "IE_MCP_ALLOWED_ORIGINS": "http://legacy01.local,http://legacy02.local",
        "IE_MCP_TIMEOUT_MS": "10000"
      }
    }
  }
}

• Escape backslashes in paths within JSON (C:\\...).

• Specify the absolute path of the built dist/index.js in args.

• For Claude Code, it can also be registered with claude mcp add.

claude mcp add ie-mode --env IE_MCP_DRIVER_PATH=C:\tools\IEDriverServer.exe --env IE_MCP_ALLOWED_ORIGINS=http://legacy01.local -- node C:\ie-mode-mcp\dist\index.js

After registration, if the client shows 10 Tools including browser_start, the connection is successful.

8. Tool Reference

There are 10 Tools exposed. Low-level WebDriver APIs (such as findElement / executeScript) are not exposed.

Common: Selector

{ "by": "id | name | css | xpath | linkText", "value": "searchButton" }

Legacy web applications frequently use name and xpath, so they are supported.

Common: frame (iframe is one level)

All element operation Tools accept an optional frame. When specified, they switch back to defaultContent and then to the frame, and search for elements within it.

{
  "frame": { "by": "name", "value": "mainFrame" },
  "selector": { "by": "id", "value": "searchButton" }
}

browser_start

{}

{ "status": "ready", "reused": false }

reused: true indicates that the existing session was used as-is. If the existing session is dead, it automatically restarts.

navigate

{ "url": "http://legacy01.local/customer" }

{ "url": "http://legacy01.local/customer", "title": "顧客検索" }

inspect_page

Main Tool for the Agent to understand the screen. Does not return full HTML, only returns URL / title / displayed text / operable elements (a button input textarea select iframe). Hidden elements and type="hidden" inputs are excluded.

{ "frame": { "by": "name", "value": "mainFrame" } }

{
  "url": "http://legacy01.local/customer",
  "title": "顧客検索",
  "text": "顧客検索 顧客名 支店 検索",
  "elements": [
    { "tag": "input", "id": "customerName", "name": "customerName", "type": "text" },
    { "tag": "select", "id": "branch", "name": "branch", "text": "東京支店", "optionCount": 12 },
    { "tag": "button", "id": "searchButton", "text": "検索" },
    { "tag": "iframe", "name": "mainFrame" }
  ],
  "truncated": false
}

• truncated: true indicates that elements were truncated at the upper limit (300 items).

• If the element list includes iframe, call again with frame specified to see its contents.

click

{ "selector": { "by": "id", "value": "searchButton" } }

{ "url": "http://legacy01.local/customer", "title": "顧客検索" }

Waits until visible and enabled, then clicks. click does not automatically retry (to prevent duplicate processing from re-clicking when registration/update/submission has already succeeded).

type

{
  "selector": { "by": "id", "value": "customerName" },
  "text": "山田太郎",
  "clear": true
}

If clear (default true) is true, inputs after clear(); if false, appends.

select

{
  "selector": { "by": "id", "value": "branch" },
  "by": "text",
  "value": "東京支店"
}

{ "text": "東京支店", "value": "13", "index": 2 }

by is one of text / value / index (index is 0-based).

wait_for

Does not use fixed sleep, waits explicitly.

{
  "type": "visible",
  "selector": { "by": "id", "value": "resultTable" },
  "timeoutMs": 10000
}

When timeoutMs is omitted, IE_MCP_TIMEOUT_MS is used.

switch_window

{ "target": "newest" }

{ "index": 1 }

{ "url": "http://legacy01.local/detail", "title": "顧客詳細", "index": 1, "windowCount": 2 }

newest polls briefly until a new window handle appears. If not detected, switches to the last existing window.

screenshot

{}

Returns a PNG image (MCP image content). Used to check layout/error screens that cannot be determined from DOM alone.

9. Usage Examples

Basic Loop

browser_start → navigate → inspect_page → click / type / select → wait_for → inspect_page

Repeat: inspect_page to understand the screen → operate → wait_for to wait for result → inspect_page again.

Example: Search for customer "Yamada Taro" and open detail screen

Example: Operate inside an iframe

{"tool": "inspect_page", "args": {}}
{"tool": "inspect_page", "args": { "frame": { "by": "name", "value": "mainFrame" } }}
{"tool": "click", "args": {
  "frame": { "by": "name", "value": "mainFrame" },
  "selector": { "by": "id", "value": "searchButton" }
}}

Specify the frame each time for each operation (because internally it resets to defaultContent and switches each time, state is not carried over).

Example: Operate a popup and return to the original window

{"tool": "click",         "args": { "selector": { "by": "id", "value": "openPopup" } }}
{"tool": "switch_window", "args": { "target": "newest" }}
{"tool": "inspect_page",  "args": {}}
{"tool": "switch_window", "args": { "index": 0 }}

10. Errors and Troubleshooting

Errors are returned with the following code instead of Selenium's stack trace (isError: true).

{
  "error": "ELEMENT_NOT_FOUND",
  "message": "Element was not found: id=searchButton",
  "selector": { "by": "id", "value": "searchButton" }
}

Recovery from DRIVER_LOST

If the browser or Driver crashes, the internal WebDriver is discarded and subsequent operations will result in BROWSER_NOT_STARTED. Automatic recovery and automatic re-execution of the last operation are not performed (to prevent side effects such as duplicate registration). The Agent should call browser_start again, check the screen state with inspect_page, and then resume operations. Since the previous operation may have already succeeded, do not re-execute registration/update operations as-is.

11. Logging

Since stdout is used by the MCP protocol, all logs are output to stderr as single-line JSON.

{"level":"info","event":"started","transport":"stdio"}
{"level":"info","tool":"navigate","url":"http://legacy01.local/customer","durationMs":842}
{"level":"info","tool":"type","selector":{"by":"id","value":"password"},"textLength":16,"durationMs":128}
{"level":"error","tool":"click","selector":{"by":"id","value":"x"},"error":"ELEMENT_NOT_FOUND","message":"Element was not found: id=x","durationMs":5012}

The input string itself, cookies, authentication information, and full HTML are not recorded (type only records character count). To save to a file, redirect stderr.

node dist/index.js 2>> C:\logs\ie-mode-mcp.log

12. Troubleshooting

screenshot is effective for investigating causes. It allows you to check states that cannot be determined from DOM information alone (modals, authentication dialogs, rendering issues).

13. Development

src/
├─ index.ts      MCP Server のエントリーポイント（stdio）
├─ config.ts     環境変数と stderr ログ
├─ tools.ts      MCP Tool の Schema と Handler
├─ browser.ts    BrowserManager（Selenium / IEDriver 操作の集約）
├─ selectors.ts  Selector → Selenium の By 変換
└─ errors.ts     Selenium Error → MCP Error Code 変換

npm run build   # tsc でビルド
npm start       # node dist/index.js

• MCP Tool does not directly touch Selenium; it always goes through BrowserManager.

• All WebDriver operations are serialized via a Promise Chain, so even if Tools are called in parallel, only one request at a time is sent to IEDriver.

• Only operations without side effects (element search, Window Handle detection) are retried. click and submissions are not retried.

14. Limitations

The initial implementation does not support the following:

Multiple browser sessions / Multiple users / HTTP Transport / REST API / DB / Session persistence / Automatic browser recovery / Complex Retry Policy / WebDriver Grid / General-purpose Selenium API / executeScript Tool / Multi-level iframes (only one level) / Element Cache / Metrics / Approval flow / Authentication and authorization