Desktop Commander MCP
Give your AI agent full control of your computer: persistent shells, background processes, filesystem reach, structured documents, and large-scale search.
Evidence: The skill explicitly requires user confirmation before high-risk SSH/deploy commands and suggests previewing destructive operations. The MCP provides allowedDirectories, command blocklist, symlink protection, and SECURITY.md clearly states these are guardrails not a sandbox, recommending Docker isolation. However, the skill does not mandate confirmation for all sensitive operations, and this static review cannot verify runtime permission enforcement. Deduction: confirmation mechanisms are only mentioned in examples, not comprehensively enforced; data-flow transparency (e.g., telemetry) is not disclosed in the skill. Full marks require complete least privilege and explicit confirmation.
Evidence: The skill provides multiple concrete workflow examples, and the repo includes integration tests (e.g., edit-block-performance.js) covering key paths, but this is a static read and cannot execute verification. Capped at 10. Deduction: tests not actually run; the skill is descriptive and depends on MCP tools, lacking self-contained executables or self-tests.
Evidence: The skill names clear scenarios (debugging, deployment, data analysis) with tool sequences, and trigger timing is clear. However, the audience is an agent, not a human, risking false triggers. Environment fit: no mention of mainland-China network reachability; core functions like SSH to overseas servers may be affected but not entirely dependent. Deduction: boundaries for non-fit scenarios not systematic; Chinese support absent (docs all English).
Evidence: Skill has clear name/description, version 0.1.0, install notes in README, MIT license, maintenance responsibility (author Eduard Ruzga). However, the skill lacks detailed parameter documentation (deferring to MCP tools), and no FAQ or known-limitations section. Deduction: information architecture is layered, but hidden assumptions (e.g., absolute paths) not fully explained; versioning and changelog incomplete.
Evidence: The skill provides workflow examples, clear goals and tool sequences, and the repo has tests suggesting task completion. But static review cannot verify actual output usability, and the skill is only a guide; results depend on MCP implementation. Capped at 7. Deduction: no runtime verification; marginal value over manual work not quantified.
Evidence: The repo includes a real CI workflow (codespell.yml) and multiple test files (ab-test.test.js, edit-block-performance.js) covering MCP key paths, but the skill itself has no verification material, and static reading was not executed. Capped at 5. Deduction: tests cover the MCP server, not the skill document itself, and no independent reproduction.
- This skill grants AI terminal control and can execute arbitrary commands; always run in an isolated environment (e.g., Docker) and only connect trusted AI clients.
- The skill requests user confirmation before high-risk operations, but this is not enforced; users should supervise critical actions.
- Documentation is in English only, no Chinese support; mainland-China users may face network restrictions when accessing overseas services (e.g., SSH).
- This review is static only; tests were not executed, so actual reliability requires further verification.
What does this skill do, and when should you use it?
This skill enables AI agents like Claude to interact with the user's machine through the Desktop Commander MCP server. It provides persistent shell sessions, long-running background processes, file access beyond the IDE sandbox, structured editing of Excel, PDF, and Word files, and ripgrep-backed search. It also supports SSH connections to remote machines and process management. The skill acts as an overview and guide for using the MCP's tools effectively in common workflows, not a replacement for the MCP itself.
The skill allows AI to: - Start and maintain persistent shell sessions (e.g., SSH, Python REPL, Node REPL) with state carried across calls; - Launch background processes (like dev servers) and get a process handle for output reading, interaction, and termination; - Read, write, move, list files within user-allowed directories, supporting files and directories; - Perform surgical text replacements via edit_block, with Excel range updates; - Handle Excel (via JSON ranges), DOCX (via raw XML edits), and PDF (create, insert/delete pages); - Use streaming ripgrep search for filename/content patterns with pagination and multiple concurrent searches; - Manage remote machines via persistent SSH sessions for diagnostics and deployment; - List, inspect, tail, and kill OS processes; - Provide session recovery via get_recent_tool_calls, list_sessions, etc.
- A developer debugging a production issue: connects via SSH, tails logs, runs diagnostics in the same session.
- An ops engineer deploying to staging: starts a deploy command, monitors output, sends interactive confirmations as needed.
- A full-stack developer iterating on an API: starts a dev server in the background, edits code, sees reload output directly.
- A developer refactoring across a monorepo: searches for all call sites, reads files in context, edits precisely, verifies zero remaining hits.
- An analyst updating Q3 numbers in Excel and tweaking a DOCX report: reads cell ranges, updates cells, edits raw XML fragments based on outline.
- A data scientist analyzing a 200MB CSV: loads it once into a Python REPL, reuses the session for multiple queries and plots.
What are this skill's strengths and limitations?
- Persistent shell sessions eliminate repeated environment setup
- Supports background processes and process management for dev servers, builds, etc.
- Directly handles Excel, Word, and PDF with near-native editing
- Ripgrep-based search scales to large projects
- SSH support extends ops capabilities to remote machines
- Session recovery tools like `get_recent_tool_calls` for context
- File operations limited to `allowedDirectories`; out-of-scope paths are denied
- `edit_block` safety depends on correct `old_string`; errors can corrupt content
- Terminal commands are not restricted by `allowedDirectories`, posing security risks
- No explicit testing on Windows/Linux mentioned; cross-platform behavior may vary
- Requires Node.js (unless using Docker installation)
How do you install this skill?
This skill is part of the Desktop Commander MCP collection. Install the entire MCP server by running npx @wonderwhy-er/desktop-commander@latest setup (requires Node.js), or use the bash installer (curl -fsSL https://raw.githubusercontent.com/wonderwhy-er/DesktopCommanderMCP/refs/heads/main/install.sh | bash). Manual configuration for Claude Desktop is also possible (see README). The skill file resides in the repository at plugins/claude/skills/desktop-commander-overview/.
How do you use this skill?
Ensure the MCP server is installed and configured. Then, give the AI natural-language commands like "Debug this production issue" or "Update the Q3 numbers in this spreadsheet." The skill guides the AI to use the correct tool sequences, such as start_process to launch a session, interact_with_process to send commands, read_file and edit_block to manipulate files. Detailed tool parameters are defined in the MCP itself.