Flower Control
Allows automated control of webpages through the Brave browser via the Web channel. It uses Playwright to read page text, links, buttons, and form fields, then perform actions such as filling forms, clicking elements, managing tabs, and uploading or downloading files. Brave is currently the only supported browser for this Web integration.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Flower Control
Windows 11 x64 source preview · 0.1.0-preview.3
Flower Control gives Codex three local MCP channels for operating Windows and the web:
Channel | What it does | Good for |
Web | Reads page structure and acts on page elements | Forms, search, web editors, uploads/downloads, temporary browsers |
App | Reads controls exposed by Windows applications | Buttons, fields, lists, menus, settings pages, dialogs |
Computer | Captures one selected window and sends pointer/keyboard input | Canvas, drag and drop, shortcuts, interfaces with weak accessibility |
Codex plans the task. Flower handles local target binding, permission checks, input release, and result readback. You do not have to call every tool by hand; describe the task in normal language.
Read these first
Install: Windows 11, Codex desktop, CPython 3.14, .NET 10 SDK, and Visual Studio C++ Build Tools.
Permissions and settings: the four tiers are 01 fully stopped, 02 ordinary control, 03 administrator windows, and 04 administrator execution.
Usage: Web, App, Computer, target selection, pause, Stop, and common recovery paths.
Setting up another PC? Use the clone commands and Codex installation prompt. Already installed? Follow first use: choose 02, try a blank Notepad document and public page, then finish control.
This is a source preview. The repository contains source and a reproducible local build; it does not contain a signed executable and does not auto-update. A GitHub Release is a version page with source archives and checksums, not an installer and not a portable helper.
Related MCP server: OpenSky
What is included
Managed Brave, Edge, and Chrome profiles for Web, with temporary, AI, and personal login profiles.
App control through Windows UI Automation, including buttons, fields, lists, menus, screenshots, and short recordings.
Computer control through a selected-window screenshot, pointer/keyboard input, drag, scroll, keys, and bounded short sequences.
Tray settings, permission tiers, deny lists, privacy masks, Stop, and Computer pause/resume.
A local Codex plugin integration. Source adapters for DSH, Claude, and local clients are retained, but are not the primary acceptance path for this preview.
Permission tiers
Tier | Name | Ordinary windows | Administrator windows | Administrator launch | Notes |
01 | Fully stopped | No | No | No | Stops all three channels; tray and Stop remain available |
02 | Ordinary control | Yes | No | No | Everyday windows without administrator permission |
03 | Administrator windows | Yes | Yes | No | Read/control elevated windows, but do not start an elevated process |
04 | Administrator execution | Yes | Yes | Yes, through the local list | Starts administrator programs through the local list; UAC secure desktop and the legacy system takeover are frozen |
Permissions are local settings. The chat agent cannot write the protected permission file. Deny lists win over allow lists. Every target, launch, and action is checked again against the current tier and Stop state.
First trial
After installation and Hook trust, choose actual tier 02 Ordinary control in tray Settings and wait for the saved tier to update. Tier changes apply immediately; ordinary list drafts use Ctrl+S. Open a new Codex chat and a blank Notepad document, then send:
Check Flower's three channels. List Notepad windows so I can identify my new blank document, bind it, type “Flower test passed”, and read it back. Do not save or close it. Finish control afterwards, preserving the text.
Open https://example.com in a temporary Flower browser, read its title, then close only that temporary session.
Use a test window or a public page first. Do not give the agent passwords, verification codes, payment information, or private browser profiles.
Browser profiles and privacy
Temporary browsers are disposable. The AI profile is an independent profile for ordinary agent work. The personal profile, Luohua, requires an explicit per-chat grant and a separate grant for each browser. Flower does not import your everyday browser cookies. See browser profiles.
Stop, pause, and recovery
Ctrl+Alt+9 ends the current Computer input and releases held keys. It does not change the saved tier or create a global all-writes Stop.
HUD Stop stops the current Computer cycle.
Tray Stop Computer Use ends the current computer operation; App/Web do not inherit a new permission from that Stop.
Computer pause stops Computer mouse/keyboard input while App/Web can continue.
Resume requires a fresh observation. Never replay an uncertain action automatically.
Privacy and security
Flower does not add telemetry, crash uploads, remote configuration, or automatic updates. Browsing reaches the sites you request; Codex still uses its configured model service. Jev, Perplexity, and compatible decision providers are optional external services and only run after local enablement. Keys enter locally in Settings, never through chat or command-line arguments.
The administrator helper runs under the current Windows account. It does not disable Defender, the firewall, or UAC.
Current limits
Web is not arbitrary browser automation; it uses supported Chromium-family browsers and managed profiles.
Computer is not desktop-wide control; it operates the bound window.
UAC secure desktop, password entry, cross-user elevation, silent updates, and indefinite key holds are outside this preview's promise.
Permission and source changes are rechecked before later work; uncertain actions are not replayed automatically.
Release status
The B46-r6 line has an installation, a local Codex publication, and three-channel load evidence on the development machine. The source archive records a per-file SHA-256 manifest. A clean-PC one-click installation, every website, every third-party app, and every administrator scenario are not implied by that record. For an issue, include the version, status, error code, and a minimal reproduction; see troubleshooting.
License
Flower's own code is under the MIT License. MCP, Playwright, FlaUI, Pillow, and other dependencies retain their licenses; see third-party notices and dependency inventory.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Let your AI agent operate any website through a real, logged-in browser session.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Connect AI assistants to your approved computer for files, commands, browser and desktop tasks.
AI-powered browser automation — navigate, click, fill forms, and extract data from any website.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables Windows agents to control desktop applications and isolated Chromium sessions through a unified computer tool, supporting discovery, observation, input, window management, browser automation, and sequencing actions.551 npm2MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to automate desktop operating systems and applications through UI automation, screenshots, and input simulation.1MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to automate both web browsers and native desktop applications through a unified interface, supporting tasks like browsing, clicking, typing, and file clipboard operations via exposed MCP tools.1Apache 2.0
- AlicenseAqualityBmaintenanceEnables agents to perceive and control a Windows desktop, a connected local Chromium browser, or a cloud Playwright browser through human-like mouse moves, clicks, drags, scrolling, and screenshots, forming an observe-locate-act computer-use loop. Includes safety guards such as random pauses, emergency stop, and approval prompts for destructive actions.11MIT