Skip to main content
Glama
io.github.flujo-app/mcp-virtual-computer

io.github.flujo-app/mcp-virtual-computer

Official
by flujo-app

MCP Virtual Computer

A VM for your agent. And for you.

  • click the mug to switch between a virtual desktop or a real one

  • unplug the network cable and cut network access

  • use the computer screen like a normal screen

  • copy and paste text between the host browser and the Xfce desktop

  • model can use terminal, create/read/edit files, see the screen, click, type

  • little fun thing: if either you or the llm types text, you see that on the keyboard; if one uses the cursor, you see the mouse move on the table.

  • automatically installs Docker Desktop on Windows for local mode, or flyctl for Fly mode

MCP client example

Local Docker (default)

COMPUTER_ID defaults to agent-workstation when omitted. Set it explicitly to keep using a named computer or to create a separate one. On a fresh Windows PC, setup prepares WSL 2 before installing Docker Desktop. Approve the Windows administrator prompt if shown; if a restart is required, restart Windows, reopen your MCP client, and retry. See Windows first-use validation.

{
  "mcpServers": {
    "virtual-computer": {
      "command": "uvx",
      "args": ["mcp-virtual-computer"],
      "env": {
        "COMPUTER_ID": "agent-workstation",
        "DESKTOP_ENVIRONMENT": "true",
        "NETWORK_ACCESS": "true",
        "AUTO_INSTALL_DOCKER": "true",
        "EXPOSE_LIFECYCLE_TOOLS": "false"
      }
    }
  }
}

Fly.io Machine

On a fresh machine, start the MCP once, run the exact flyctl auth login command shown by its error, and restart the MCP client

{
  "mcpServers": {
    "virtual-computer-fly": {
      "command": "uvx",
      "args": ["mcp-virtual-computer", "--backend", "fly"],
      "env": {
        "COMPUTER_ID": "agent-workstation",
        "DESKTOP_ENVIRONMENT": "true",
        "NETWORK_ACCESS": "true",
        "AUTO_INSTALL_FLYCTL": "true",
        "EXPOSE_LIFECYCLE_TOOLS": "false"
      }
    }
  }
}

Related MCP server: Docker MCP Server

What it exposes

  • terminal_execute — run a command in the configured Docker container or Fly Machine.

  • read_file — read a UTF-8 text file relative to /workspace or by absolute path.

  • write_file — atomically write a UTF-8 text file.

  • edit_file — replace one exact text match, or all matches when requested.

  • computer_ui — open or attach the Three.js computer view. Its result includes both the MCP App resource_uri and a real loopback url/dashboard_url that opens in an external browser, including for stdio clients.

DESKTOP_ENVIRONMENT defaults to true. With the real desktop enabled it additionally exposes:

  • look_at_screen — return the current PNG framebuffer, AT-SPI snapshot, or both.

  • click, type, and scroll — interact by AT-SPI element reference or coordinates.

  • list_windows and switch_window — enumerate and activate real Xfce windows.

  • move_window, maximize_window, restore_window, minimize_window, and close_window — control a selected window by ID, title, or class.

  • computer://screen/current.png — current live framebuffer resource.

  • computer://screen/accessibility.json — Playwright-like AT-SPI tree with element refs, roles, names, actions, focus state, and screen bounds.

The MCP App can always call runtime_status, set_network_access, and set_desktop_environment. Set EXPOSE_LIFECYCLE_TOOLS=true to additionally expose those lifecycle controls to the model; they remain model-hidden by default.

Architecture

Demo: FLUJO

Demo: Claude Desktop

Demo: Goose

Permanent Fly Machine

Fly mode builds the bundled Xfce image with Fly's remote builder and deploys it as one permanent Machine. The root filesystem and /workspace survive MCP restarts, and the Three.js screen uses the same noVNC desktop through Fly's HTTPS/WebSocket proxy.

This is the Xfce framebuffer returned by look_at_screen from a deployed Fly Machine. The same live desktop appears on the rendered computer after computer_ui connects to VNC:

Xfce desktop running on a permanent Fly Machine

No app name, region, CPU size, memory size, Docker installation, or VNC configuration is required. The package's MCP Registry declaration, including the docker and fly backend choices, is in server.json. See Fly setup for the complete first-run and authentication flow.

On first use, virtual-computer finds fly/flyctl or downloads the current official release to ~/.fly/bin. It uses a cached fly auth login session or FLY_API_TOKEN, selects the personal organization when available, creates and remembers a generated app, lets Fly choose the closest placement, and defaults to one shared CPU with 1 GB RAM.

Authentication is the only unavoidable account step. On a fresh machine, start the MCP once, run the exact flyctl auth login command shown by its error, and restart the MCP client. The first computer call can take several minutes while Fly remotely builds Xfce. Later starts attach to the same Machine.

Optional overrides remain available as FLY_ORG, FLY_APP_NAME, FLY_REGION, FLY_API_TOKEN, and the --fly-* flags. Set AUTO_INSTALL_FLYCTL=false to require a preinstalled CLI. Deleting or factory-resetting the computer destroys its persistent Fly root filesystem; ordinary MCP shutdown leaves the permanent Machine running. Stop or delete it from Fly when you no longer want it to incur usage.

License and origin

MIT licensed. This project is a persistent-computer fork of Kilntainers with Docker and Fly backends.

Available Tools

9 tools
computer_uiOpen Virtual ComputerB

Open the interactive Three.js laptop and its computer screen.

Inputs: no named parameters.

edit_fileB

Replace exact UTF-8 text and show it being selected, overwritten, and saved.

Inputs: path (required), new_text (required), old_text (required), replace_all.

list_directoryB

List a Docker directory for the interactive virtual Explorer.

Inputs: path.

read_fileB

Read a UTF-8 text file and show it being opened and scrolled on the virtual computer.

Inputs: path (required).

runtime_statusC

Read genuine lazy container-runtime setup progress.

Inputs: no named parameters.

set_desktop_environmentB

Switch the running computer between virtual and Xfce desktops.

Inputs: enabled (required).

set_network_accessC

Change the running Docker computer's real network access.

Inputs: enabled (required).

terminal_executeA

Execute a shell command in one persistent Debian Docker computer. Commands run in bash. Each call is independent — no state (shell variables, working directory) persists between calls (however filesystem does persist). Use the working_directory parameter or chain commands with && to control execution context.

To write files or pass data without shell escaping, use the stdin parameter (e.g., command="cat > file.txt" with content in stdin). Commands time out after 120 seconds by default (override with the timeout parameter for long-running operations).

Inputs: args, stdin, command, timeout, +1 more.

write_fileA

Write a UTF-8 text file and show it being typed and saved on the virtual computer.

Inputs: path (required), content (required), create_parent_directories.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 9 tools

    Disambiguation4/5

    Most tools target clearly distinct actions: shell execution, file read/write/edit/list, network, desktop, UI, and status. Some overlap exists between terminal_execute and the dedicated file tools, since terminal commands can also create and edit files, but the descriptions clarify their intended specializations.

    Naming Consistency3/5

    The majority follow a verb_noun pattern (read_file, write_file, list_directory, edit_file, set_network_access, set_desktop_environment), but terminal_execute, computer_ui, and runtime_status break that pattern with noun-first or noun-only names. The inconsistency is noticeable but not chaotic.

    Tool Count5/5

    Nine tools is a well-scoped set for a virtual computer MCP server. Each tool addresses a distinct part of the workflow: terminal access, file operations, runtime status, UI launch, and configuration switches, with no obvious bloat.

    Completeness4/5

    The tool surface covers the main lifecycle: executing commands, reading/writing/editing/listing files, controlling network and desktop environment, and checking runtime status. Missing file operations like delete or rename are not critical because terminal_execute can handle them, but they are minor gaps in the dedicated file-tool set.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Provides isolated Docker environments for code execution, enabling users to create and manage containers, execute multi-language code, save and reproduce development environments, ensuring security and isolation.
      16
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables secure execution of shell commands and file operations within isolated Docker containers. Provides process management, interactive input handling, and comprehensive file system operations for containerized development environments.
      128 npm
      6
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Enables code execution in isolated Docker containers with persistent IPython, Node.js, or R kernels, supporting file import/export and cross-session transfers via MCP tools.
      6
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to operate a real, private, local Docker-based computer with durable files, terminal access, web research, and a persistent browser or desktop. It supports multiple agent clients via MCP or OpenAPI, with live viewing and human takeover.
      13
      Apache 2.0