SkillAgentSearch skills...

cmdshellmcp

command shell mcp server with additional tools (e.g. for coding)

Install / Use

claude mcp add ag88 -- npx -y github:ag88/cmdshellmcp

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

80/100

Supported Platforms

Claude Code
Claude Desktop

Tags

Our assessment of cmdshellmcp

cmdshellmcp scores 80/100 on our quality scale, 746th of 957 AI & Machine Learning skills we index.

Its MCP Server is 34 KB long, well organised into 40 sections with 30 code examples: a thorough specification that gives an agent plenty to work with.

It has 3 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
12/15
Adoption
3/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated 31 days ago, so cmdshellmcp is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 87/100, with 2 cautions from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-11-16. Automated pattern scan on 2026-11-16. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

cmdshellmcp compared with similar skills

All 4 of these similar skills score higher than cmdshellmcp; compare them before choosing.

SkillScoreStarsUpdatedFormat
cmdshellmcp (this skill)by ag8880331d agoMCP Server
claude-memby thedotmack100117.7ktodayCLAUDE.md
Agent-Reachby Panniantong100112.4k9d agoCLAUDE.md
Understand-Anythingby Egonex-AI10092.9k1d agoCLAUDE.md
rufloby ruvnet10081.9ktodayCLAUDE.md

Frequently asked questions

How do I install cmdshellmcp?
Run claude mcp add ag88 -- npx -y github:ag88/cmdshellmcp. The install tabs above show the steps for each supported agent.
Which AI agents does cmdshellmcp work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is cmdshellmcp safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 87/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is cmdshellmcp still maintained?
The repository was last updated 31 days ago, so cmdshellmcp is actively maintained.

cmdshellmcp

cmdshellmcp is a constrained command-shell MCP server for AI agents. It exposes a small allowlisted Unix command set, somewhat safe file operations, patch application, and URL fetching so an MCP client can perform limited local tasks without unrestricted shell access.

The server is implemented in Python and runs as an MCP server using the fastmcp package. By default it listens on 127.0.0.1:8003 using the streamable HTTP transport unless --sse is selected.

Added setup files for a full featured Docker Debian based sandbox system container with cmdshellmcp (setup scripts, guides DOCKER.md, Dockerfile, cmdshellmcp.service unit file that runs cmdshellmcp in the container).

[!CAUTION]

This server provides remote command execution (RCE), which is normally considered a critical security vulnerability. Allow-listing powerful commands—such as bash, sh, python, perl, sudo, docker, or commands capable of writing files—may allow an attacker or untrusted LLM to bypass intended restrictions and take control of the system. For example, allowing python or bash can effectively permit arbitrary code execution and file access. If the server lacks strong authentication, is reachable by untrusted clients, or is controlled by an untrusted or prompt-injected LLM, it may cause severe damage, including data loss, credential theft, malware installation, or compromise of other systems. Run cmdshellmcp (cmdshellmcp2.py) only in a protected, disposable sandbox with carefully scoped privileges limited to those required for its intended task and limited access to files, credentials, devices, and networks—for example, disposable Docker container or virtual machine that can be safely destroyed after use.

Warning

This is a scratch / experimental app that evolved out of using an MCP server (tools) that allow running shell commands for coding purpose. Unfortunately, for such a purpose, one often needs to provide the LLM (model) client with rather powerful shell commands to "do its job" for a particular scope / context / intent.

The default allowed commands are not necessarily safe, i.e. LLM agents or practically clients calling the MCP api can 'escape' and do things outside a context e.g. the working directory defined with --cwd option. It validates the arguments in a limited way, i.e. looking for string path patterns if they seem to escape the current directory. This is a heuristic check, i.e. this check is pattern matching, it does not know if that argument is intended to be a file or directory. Nor does it know the argument's context for a particular command, or if they are after all safe. e.g. it may prevent a ls / (listing a root directory), but sometimes that is what you need. But it may prevent by sheer pattern matching e.g. rm /PreciousFile.txt i.e. remove a file in the root directory. This is not an always secure check, but added as it is better than no checks at all.

There are also tools (MCP functions exposed) that expose write and file modification ops, including executing shell commands.

  • Authentication is enabled by default. Keep the randomly generated token, or set a fixed auth_token using the auth field in the config file or --auth. Using --noauth practically means you are giving remote command execution (RCE) to any (including possibly malicious or rogue) clients that can reach the server.
  • Run this as an unprivileged user. Running as root is at best foolish
  • Do not use this with untrusted clients or untrusted LLMs
  • Use it in a disposable sandbox e.g. a standalone docker container or virtual machine that you can afford to throw away including the contents
  • Review the allow list in cmdshellmcp.json and the hardcoded defaults, revise them before using.

Features

  • Allowlisted shell execution for a curated set of commands
  • File read/write/list/copy/move/rename/delete operations and basic directory creation/removal under a configured working directory
  • Transactional text-file editing with sed, numbered backups, and unified diffs
  • Unified diff patch application via patch
  • HTTP fetch support with optional HTML prettification
  • Bearer token authentication with a secure, randomly generated token by default
  • Audit logging to stdout and/or a file
  • Path restrictions to prevent escaping the current working directory
  • Setup files for a delightful, full featured Docker Debian based sandbox system container with cmdshellmcp (setup scripts, guides DOCKER.md, Dockerfile, cmdshellmcp.service unit file that runs cmdshellmcp in the container).

AI use in this repo

This app and its contents e.g. this page, is created with the aid of LLM (large language models) such as

  • ChatGPT 5.6 sol (light), Codex
  • Github Co-pilot MAI-Code-1.1-Flash

Initial codes is written by the author and refactored with aid of the LLMs and updates is also done partly manually. After features are added or changed, additional tests is often done by manually running them e.g. in llama.cpp's llama-server Web UI

Installation

  1. Clone the repository.
  2. Create and activate a virtual environment if desired.
  3. Install dependencies:
python -m venv .venv
source .venv/bin/activate
pip install fastmcp requests beautifulsoup4

If your environment uses a requirements.txt, you can also install from there:

pip install -r requirements.txt

The editFile tool also requires GNU sed (including its --sandbox option) and diff to be installed on the server. These programs are invoked directly by the dedicated tool and do not need to appear in allowed_commands.

Configuration

A configuration file is mandatory. The server selects one JSON file in this order (files are not merged):

  1. The file specified by --conf, when given (no automatic search).
  2. /etc/cmdshellmcp.d/cmdshellmcp.json.
  3. /usr/local/python/cmdshellmcp/cmdshellmcp.json.

The current directory is never searched automatically. If no config is found, startup exits with an error suggesting --conf FILE. A missing explicit file, unreadable file, or invalid selected JSON file also fails startup. Command-line options override their configuration keys; omitted optional settings retain internal defaults within the required configuration.

Example:

{
  "host": "127.0.0.1",
  "port": 8003,
  "cwd": "/path/to/project",
  "shellnopathchk": false,
  "editdelbk": false,
  "quiet": false,
  "auditlog": null,
  "disableTools": ["writeFile", "editFile", "applyPatch", "copyFile", "moveRenameFile", "deleteFile", "mkdir", "rmdir"],
  "allowed_commands": [
    "ls", "pwd", "date", "cat", "grep", "egrep",
    "whoami", "head", "tail", "sed", "wc", "file", "du", "df",
    "free", "ps", "uname", "hostname", "uptime", "w", "last",
    "mkdir", "cp", "mv", "awk"
  ],
  "auth": "_my_secret_auth_token_"
}

Supported configuration keys:

  • host: bind host; default 127.0.0.1
  • port: bind port; default 8003
  • cwd: working directory path; required unless --cwd is supplied
  • shellnopathchk: boolean; bypasses shell argument path checks when true; default false
  • editdelbk: boolean; deletes successful edit backups after composing the diff when true; default false
  • quiet: suppresses audit output to stdout when true
  • auditlog: optional path to an audit log file
  • allowed_commands: list of commands permitted for execution
  • disableTools: list of MCP tool names to omit from the server, case sensitive and exact name match is required
  • auth: optional fixed bearer token string; when omitted, a random token is generated at startup

Running the server

Start with an explicit configuration and working directory:

python cmdshellmcp2.py --conf /path/to/cmdshellmcp.json --cwd /path/to/project

If the config is installed in a default search location and supplies cwd, python cmdshellmcp2.py is sufficient. Startup fails when neither --cwd nor config cwd supplies a directory. There is no prompt or automatic working directory fallback. The selected directory is logged before the server starts. By default, the server listens on 127.0.0.1:8003 using streamable HTTP. The following examples assume a config is installed in a default search location; otherwise add --conf FILE. Supply --cwd PATH if the config leaves cwd unset.

SSE mode

python cmdshellmcp2.py --sse

This runs the server on the SSE transport instead of streamable HTTP.

Browser clients and CORS

Both HTTP transports include CORS middleware for browser-based MCP clients. The server accepts requests from any origin, supports the MCP GET, POST, and DELETE methods and browser preflight OPTIONS requests, and exposes the mcp-session-id response header to browser JavaScript.

Because all origins are allowed, do not expose the server to an untrusted network without authentication and appropriate network controls. To restrict browser access, replace allow_origins=["*"] in CORS_MIDDLEWARE with the specific trusted origins.

Custom working directory

python cmdshellmcp2.py --cwd /path/to/project

This overrides config cwd for file and shell tools. A configuration file is still required. The path must exist and must be a directory; ~ is expanded and the selected path is normalized to an absolute path.

Custom host and port

python cmdshellmcp2.py --host 0.0.0.0 --port 9000

Authentication

Authentication is enabled by default. If neither --auth nor an auth value in the configuration file supplies a fixed token, the server generates a cryptographically secure, base64-encoded token and displays it in the startup log. The generated token is shown in bold when stdout supports ANSI styling. Configure the MCP client to send that value as its bearer token.

To use a fixed token instead:

python cmdshellmcp2.py --auth my-secret-token

To explicitly run without authentication:

python cmdshellmcp2.py --noauth

--auth and --noauth are mutually exclusive. --noauth also overrides an auth value in the configuration file. Disabling authentication is unsafe on any network that is not completely trusted.

Audit logging

python cmdshellmcp2.py --quiet --auditlog /tmp/cmdshellmcp.log
  • --quiet disables audit logging to stdout.
  • --auditlog appends audit events to the specified file.

A full featured Docker Debian based sandbox system container with cmdshellmcp

This project provides setup files for a delightful, full featured Docker Debian based sandbox system container with cmdshellmcp (setup scripts, guides DOCKER.md, Dockerfile, cmdshellmcp.service unit file that runs cmdshellmcp in the container).

Using the docker sandbox container gives you a "system in a system" experience, e.g. that Debian boots up within the container just like a full Unix / Linux system. It gives you 3 userids within the container:

  • root: (superuser)
  • admin (this is for you / human user), admin has sudo access, i.e. practically can do anything root does, install files, change system configs, permissions etc
  • codeagent: cmdshellmcp runs under this user, in the user's /home/codeagent home directory. codeagent has no sudo access.

For more details review DOCKER.md

Command-line options

python cmdshellmcp2.py [--cwd PATH] [--host HOST] [--port PORT] \
  [--allow COMMAND [COMMAND ...]] [--conf FILE] [--auth TOKEN | --noauth] \
  [--disableTools TOOL[,TOOL...]] [--editdelbk] [--shellnopathchk] [--sse] [--quiet] \
  [--auditlog FILE] [--stdio]

Options:

  • --cwd: working directory used for file and shell operations; when omitted, use config cwd, otherwise exit with a missing working directory error
  • --host: server bind host
  • --port: server bind port
  • `--

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars3
CategoryAI
Updated1mo ago
Forks0

Languages

Python

Trust signals

87/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

2 low