Shell

MCP Shell Server

A trusted execution server implementing the Model Context Protocol (MCP). It runs allowlisted commands as argv arrays, with stdin input, contained redirection, a minimal child environment, execution limits, and structured audit logging.

This package is not a sandbox. Allowing a command delegates that program the server process's existing OS authority. Read Trusted execution contract before configuring it.

Features

• Argv-based Command Execution: Allowed commands run via subprocess argv without shell-string interpretation
• Standard Input Support: Pass input to commands via stdin
• Comprehensive Output: Returns stdout, stderr, exit status, and execution time
• Safe Pipeline Support: Pipelines preserve and validate argv segments instead of invoking a shell
• Execution Limits: Server-side default timeout, maximum timeout, and output byte caps are enforced
• Contained Redirection: <, >, and >> targets must stay inside the requested working directory
• Minimal Child Environment: Child processes receive a small allowlisted environment instead of inheriting all server secrets
• Structured Audit Logging: Success, rejection, timeout, output-cap, and process-error outcomes are logged with redaction

Trusted execution contract

Read this before the configuration examples below. The server emits the same warning once at startup through its logger (stderr); MCP stdout framing is unchanged.

What the server controls

mcp-shell-server enforces a boundary around its own behavior:

• which executable names it launches directly (ALLOW_COMMANDS, ALLOW_PATTERNS),
• argv-based process creation without shell-string interpretation,
• server-managed <, >, and >> redirection contained under the requested working directory,
• the child environment it builds,
• timeout and output-byte limits,
• structured audit records.

What the server does not control

Allowing a command is authority delegation: the allowed program runs with the server process's existing OS identity, filesystem access, network access, and credentials. ALLOW_COMMANDS and ALLOW_PATTERNS restrict only the command names the server launches directly. They do not guarantee containment of:

• child processes an allowed program spawns,
• program-specific interpreters, expression languages, or plugins,
• configuration files, dotfiles, or environment-driven behavior the program reads on its own,
• filesystem reads and writes the program performs itself, outside server-managed redirection,
• network access the program performs itself.

The command-specific rejection rules described under Security are best-effort defense in depth against known dangerous argument forms. They are non-exhaustive, and they are not a proof that an allowed program is safe.

Untrusted input requires external isolation

A trusted, authenticated MCP client does not make the content it processes trusted. Model-provided text, fetched web pages, issue text, and repository contents are untrusted input even when the client itself is trusted, and an LLM can be induced to construct a request from that content.

If any request or any file the server can reach may derive from untrusted input, run the server inside an independently enforced OS boundary (container, VM, jail, or OS policy) that provides:

• Least-privilege identity — a dedicated non-root user with no administrative rights on the host.
• Filesystem scope — only the directories the workload needs, mounted read-only where possible, with host configuration, SSH keys, and cloud credential files out of reach.
• Network restrictions — egress denied by default, allowing only the destinations the workload requires.
• Credential exclusion — no ambient tokens, cloud instance-metadata access, or agent sockets reachable from the server process.
• Descendant-process containment — a process/cgroup namespace so processes spawned by an allowed program are bounded and reaped with the sandbox.
• Resource limits — CPU, memory, disk, process-count, and wall-clock caps enforced outside this package.

Keep the allowlist as narrow as the workload allows; a narrow allowlist reduces exposure but never substitutes for the boundary above.

MCP client setting in your Claude.app

Published version

code ~/Library/Application\ Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "shell": {
      "command": "uvx",
      "args": [
        "mcp-shell-server"
      ],
      "env": {
        "ALLOW_COMMANDS": "ls,cat,pwd,grep,wc,touch,find"
      }
    },
  }
}

Local version

Configuration

code ~/Library/Application\ Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "shell": {
      "command": "uv",
      "args": [
        "--directory",
        ".",
        "run",
        "mcp-shell-server"
      ],
      "env": {
        "ALLOW_COMMANDS": "ls,cat,pwd,grep,wc,touch,find"
      }
    },
  }
}

Installation

Installing via Smithery

To install Shell Server for Claude Desktop automatically via Smithery:

npx -y @smithery/cli install mcp-shell-server --client claude

Manual Installation

pip install mcp-shell-server

Usage

Starting the Server

ALLOW_COMMANDS="ls,cat,echo" uvx mcp-shell-server
# Or using the alias
ALLOWED_COMMANDS="ls,cat,echo" uvx mcp-shell-server

The ALLOW_COMMANDS (or its alias ALLOWED_COMMANDS ) environment variable specifies which commands are allowed to be executed. Commands can be separated by commas with optional spaces around them.

Valid formats for ALLOW_COMMANDS or ALLOWED_COMMANDS:

ALLOW_COMMANDS="ls,cat,echo"          # Basic format
ALLOWED_COMMANDS="ls ,echo, cat"      # With spaces (using alias)
ALLOW_COMMANDS="ls,  cat  , echo"     # Multiple spaces

ALLOW_PATTERNS can be used for comma-separated regular expressions that match command names. Each pattern is applied with full-match semantics, so ALLOW_PATTERNS="ls" allows only the command name ls and does not allow lsof or ls -la. Patterns and command names containing whitespace or shell metacharacters are rejected; do not use ALLOW_PATTERNS to describe shell command strings or argument-level policies.

ALLOW_PATTERNS="python[0-9.]*,node"    # Command-name patterns only

Allowlisting a command name is not a sandbox for that program's own argument-level execution features. The server applies default argument hardening even when the binary is allowed: known exec-capable vectors such as find -exec/-execdir/-ok/-okdir, shell/interpreter launchers and JS runtimes such as deno/bun, awk system(), tar external-program options such as tar --checkpoint-action=exec (every checkpoint action is denied), env, xargs, command-wrapper tools such as timeout/nice/nohup/setarch/strace/valgrind, shell-escape tools such as sed/less/vim/ssh, common alternate names such as gfind/gawk/gtar/bsdtar/gsort, GNU sort options that select an external program, output path, external file list, or external temporary directory (--compress-program, -o/--output, --files0-from, and -T/--temporary-directory, including abbreviated and clustered forms such as --co, --o, -ro FILE, and -rT DIR), all Git command-scoped configuration overrides, persistent git config writes, and the Git subcommand execution options described below are rejected before subprocess creation. For example, ALLOW_COMMANDS="git" does not permit git -c user.name=Example status, git -c alias.pwn=!sh -c "touch marker" pwn, or git config alias.pwn '!sh -c "touch marker"'; every global git -c <name=value> and git -c<name=value> override is rejected regardless of its key or value. Dangerous names such as setarch, strace, valgrind, deno, and bun are rejected categorically, including version or attach-only forms such as deno --version or strace -p PID, and an allowlist entry cannot override this.

Git subcommand execution options are rejected by a finite Git 2.53.0 option grammar: git grep pager selection (-O, attached and clustered forms such as -OP, -nOP, -1OP, and --open-files-in-pager with any non-empty prefix such as --op), git rebase execution (-x, attached and clustered forms such as -xP, -qxP, and --exec with any non-empty prefix such as --ex, including after operands), every git filter-branch invocation, and git bisect run when run directly follows bisect. On every other subcommand, --exec and its non-empty prefixes (--e, --ex, --exe) stay rejected as remote-helper selectors (for example git push --exec=PROG). Values consumed by global options (-C, --git-dir, --work-tree, --namespace, ...) and by value-taking grep/rebase options (-e, -f, --context, --onto, -s, -X, --strategy-option, ...) are data and are not rescanned, so git grep -e -O data and git rebase --onto --exec are allowed; an unconsumed -- or --end-of-options ends grep/rebase option parsing, and grep also stops at its first operand except for (/). Unknown short letters and ambiguous prefixes that include a target (such as git grep --o=P) are conservatively rejected. git send-email program selectors are rejected by a finite Git 2.53.0 recognizer that follows its three Getopt::Long passes (identity, help/alias listing, then the main options) over the remaining arguments under both default permutation and the POSIXLY_CORRECT require-order mode; a selector reachable in either interpretation is rejected: --sendmail-cmd, --to-cmd, --cc-cmd, and --header-cmd are rejected whenever selected, including empty or missing values and with --dry-run, a later override, or --no-header-cmd; --smtp-server is rejected when its value is missing, empty, begins with / or \, or begins with an ASCII drive letter and colon (such as C:). That path rule is a conservative union of POSIX and Windows spellings that can also reject drive-relative or rooted strings Git would not run, and it is not a claim of Windows runtime isolation. Option names are case-insensitive and accepted with --, -, or + prefixes and as unique abbreviations such as --sendm; exact names such as --smtp-server-port and --smtp-server-option win, and ambiguous prefixes that include a selector (such as --send, --cc-c, --to-c, or --smtp-serv) are rejected conservatively, not because each one runs a program in Git. Ordinary SMTP hosts (--smtp-server=smtp.example.invalid, localhost, 127.0.0.1, [::1]), -h, --dump-aliases, --dry-run, values consumed by value-taking options (git send-email --subject --sendmail-cmd=DATA, --identity --sendmail-cmd=DATA), and arguments after an unconsumed -- remain allowed. Options are still recognized after operands, --end-of-options does not end send-email option parsing, and a -- consumed as a required value is data. Invalid early-pass tokens (flags with equals attachments, empty required strings, or invalid numeric values) remain residual argv. The two-mode union may reject selector-looking data that only one mode consumes. Configuration-driven helpers (sendemail.sendmailCmd, sendemail.smtpServer, sendemail.toCmd, sendemail.ccCmd, sendemail.headerCmd) and argv-driven behavior outside these selectors, such as the editor run by --compose/--annotate or options passed through to git format-patch, remain delegated authority. Compatibility: git filter-branch is rejected categorically, including git filter-branch --help. These are finite defense-in-depth rules, not complete Git confinement: hooks, editors, merge strategies, textconv drivers, and other argv- or configuration-driven helpers (including difftool commands and submodule foreach) remain delegated authority.

These rejections are a finite list of known names and argument forms; the policy does not recognize every program capable of launching helpers or other programs. Programs you allowlist keep their own delegated behavior: make recipes run their own commands, and dd/tee read and write the paths named in their own operands (of=, file arguments) with the server process's OS authority. Those program-owned operations are delegated, not contained, and are not blocked by validation. Hostile or untrusted workloads require external least-privilege OS/container isolation.

For tar and its gtar/bsdtar aliases, the options that make tar launch another program are rejected: --checkpoint-action (every action, not only exec=), --use-compress-program/-I, --to-command, --rsh-command, --rmt-command, and --info-script/--new-volume-script/-F. Rejection covers separated and = values, every abbreviation whose candidates include one of these names (such as --use-comp, --to, --rm), -I/-F anywhere in a short-option cluster or a first-position old-style word such as cIf, and options placed after operands. Arguments are checked with both a GNU tar 1.34 and a bsdtar 3.5.3/3.7.4 option grammar regardless of the executable name, and the command is rejected if either interpretation reaches a helper option; for example, bsdtar treats -H/-L as flags and -Wname as --name, where GNU tar reads them as value-taking options. Values consumed by a benign option are data, a discrete unconsumed -- ends option parsing, and ordinary create/list/extract, built-in compression switches (-z/-j/-J), and numeric --checkpoint=N intervals stay allowed. Compatibility: non-exec checkpoint actions such as --checkpoint-action=dot are now rejected, as are argument lists that are harmless in one tar dialect but run a helper in the other. This is a finite rule set, not a tar sandbox: file lists and other tar filesystem actions are not validated.

Write sorted output with the server's contained redirection instead of sort -o: ["sort", "input", ">", "output"] keeps the target inside the requested working directory, while sort -o would write directly to any process-accessible path. Option-like filenames stay usable after the -- delimiter, for example ["sort", "--", "--output=data"].

This hardening is best-effort defense in depth against known dangerous argument forms. It is non-exhaustive and is not a complete sandbox for arbitrary untrusted command execution. See Trusted execution contract for the external isolation required when requests or file contents may derive from untrusted input.

Child process environment

Commands run with an isolated child environment. The server does not pass the full parent process environment to child commands, so unrelated variables such as API tokens, credentials, and SECRET_TOKEN are absent by default.

By default the child environment contains only the minimal launch keys needed for command execution: PATH on POSIX systems, plus Windows process-launch keys when applicable (COMSPEC, PATHEXT, SYSTEMROOT, and WINDIR).

The child PATH is not inherited from the server process. It is chosen with this precedence:

1. When PATH is explicitly listed in MCP_SHELL_CHILD_ENV_ALLOWLIST, a per-call envs PATH value, otherwise the parent process PATH. Delegated values, including empty strings, are passed unchanged.
2. Otherwise a nonblank MCP_SHELL_SAFE_PATH, used exactly as configured (no trimming).
3. Otherwise, on POSIX, the default /usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin. An unset, empty, or whitespace-only MCP_SHELL_SAFE_PATH uses this default.

On POSIX, bare command names such as git are resolved using this child PATH, so an earlier executable in the server's own PATH is not selected and a command missing from the safe PATH fails instead of falling back to the parent PATH. Tools installed outside the system directories (for example Homebrew on Apple Silicon, ~/.local/bin, or version-manager shims) therefore require either adding their directory to MCP_SHELL_SAFE_PATH or explicitly delegating with MCP_SHELL_CHILD_ENV_ALLOWLIST="PATH":

MCP_SHELL_SAFE_PATH="/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin" \
ALLOW_COMMANDS="git,rg" \
uvx mcp-shell-server

Delegating PATH is an operator opt-in: without it, a per-call envs PATH value is ignored. On Windows, a nonblank MCP_SHELL_SAFE_PATH is used as the child PATH; otherwise the parent PATH is preserved (falling back to os.defpath only when absent). Windows native executable search does not follow the child environment PATH, so this setting is not a safe-lookup guarantee on Windows.

Use MCP_SHELL_CHILD_ENV_ALLOWLIST to explicitly allow additional environment variable names to be inherited from the parent process or accepted from per-command environment overrides. The allowlist is comma-separated and uses exact environment variable names:

MCP_SHELL_CHILD_ENV_ALLOWLIST="LANG,LC_ALL,MY_TOOL_HOME" \
ALLOW_COMMANDS="printenv,my-tool" \
uvx mcp-shell-server

Only keys named in MCP_SHELL_CHILD_ENV_ALLOWLIST are forwarded. Secret-like names are treated defensively in logs and should not be allowlisted unless you intentionally want a child command to read that secret.

Structured audit logs

Each command invocation emits one mcp-shell-server.audit log event named shell_execution_audit. Audit records cover successful execution, validation rejection before subprocess creation, timeout, output-cap termination, and process errors including subprocess creation failures.

Audit metadata includes:

• timestamp, duration, and result_type
• command name and redacted argv
• resolved working directory
• redirection flags for stdin/stdout/stdout append
• redacted per-call environment override metadata, when supplied
• effective timeout and output_limit
• stdout_bytes and stderr_bytes
• return_code when available
• rejection_reason or error_type where applicable

Audit logs intentionally do not include raw stdout or stderr bodies. Secret-like argv and environment names or values containing markers such as SECRET, TOKEN, PASSWORD, PASSWD, API_KEY, ACCESS_KEY, PRIVATE_KEY, KEY, CREDENTIAL, or AUTH are replaced with [REDACTED]. Long non-numeric values are represented by a short SHA-256 digest instead of the raw value.

Request Format

The directory argument is optional. If omitted, commands run in the MCP server process current working directory (server process CWD). Relative directory values are resolved from that same server process CWD. This base is not the MCP client CWD; it is the working directory of the process that launched mcp-shell-server.

# Basic command execution in the server process CWD
{
    "command": ["ls", "-l"]
}

# Command with a relative working directory resolved from the server process CWD
{
    "command": ["pwd"],
    "directory": "subproject"
}

# Command with stdin input
{
    "command": ["cat"],
    "stdin": "Hello, World!"
}

# Command with timeout
{
    "command": ["long-running-process"],
    "timeout": 30  # Maximum execution time in seconds
}

# Command with working directory and timeout
{
    "command": ["grep", "-r", "pattern"],
    "directory": "/path/to/search",
    "timeout": 60
}

Response Format

Successful response:

{
    "stdout": "command output",
    "stderr": "",
    "status": 0,
    "execution_time": 0.123
}

Error response:

{
    "error": "Command not allowed: rm",
    "status": 1,
    "stdout": "",
    "stderr": "Command not allowed: rm",
    "execution_time": 0
}

Failed single commands: when a command exits with a non-zero status, the MCP tool result has isError: true and its text starts with Error executing command: Command failed with exit status N (or Error executing command: Command terminated by signal N when the process was killed by a signal), followed by the command's stdout and stderr when present. This output is returned to the client without logging its bodies. Successful commands (exit status 0) return their stdout/stderr unchanged. Pipeline failure handling is unchanged and does not preserve output in the same way.

Security

The measures below are the server's own enforceable boundary; they are not an OS sandbox. A command-name allowlist controls which executables the server launches directly, but an allowed program still runs with the server process's OS authority and may read accessible files, consume CPU, spawn child processes, or reach the network. See Trusted execution contract for what this does and does not contain, and for the external isolation required with untrusted input.

1. Command Whitelisting: Only explicitly allowed command names or full-matching ALLOW_PATTERNS entries can be executed. This admits executable names; it does not confine an allowed program's own behavior.
2. Default Argument Hardening (non-exhaustive defense in depth): Known exec-capable vectors such as shells/interpreters, JS runtimes (deno, bun), command wrappers (setarch, strace, valgrind, timeout, ...), env, xargs, find -exec/-execdir/-ok/-okdir, awk system(), tar external-program options (--checkpoint-action, --use-compress-program/-I, --to-command, --rsh-command, --rmt-command, --info-script/--new-volume-script/-F, under both GNU and BSD tar grammar), GNU sort --compress-program/-o/--files0-from/-T, Git external-program options (including git grep -O, git rebase -x/--exec, git filter-branch, git bisect run, and git send-email --sendmail-cmd/--to-cmd/--cc-cmd/--header-cmd and program-like --smtp-server values), and every global git -c <name=value> or git -c<name=value> configuration override are rejected by default even when the command name is allowlisted. This is a finite list, not a recognizer for every program capable of launching helpers; allowlisted make recipes and dd/tee program-owned file I/O remain delegated authority.
3. No Shell-String Execution: Normal commands and pipelines are executed with asyncio.create_subprocess_exec(*argv); user-controlled strings are not passed to a shell.
4. Contained Redirection: Redirection paths must be relative to directory; absolute paths, .. traversal, and symlink escapes are rejected before files are opened.
5. Environment Isolation: Children receive a minimal environment plus names listed in MCP_SHELL_CHILD_ENV_ALLOWLIST. Parent secrets such as tokens are not inherited by default. Per-call envs values are only accepted for explicitly allowlisted names.
6. Execution Limits: MCP_SHELL_DEFAULT_TIMEOUT_SECONDS defaults to 30 seconds, MCP_SHELL_MAX_TIMEOUT_SECONDS defaults to 300 seconds, and MCP_SHELL_OUTPUT_LIMIT_BYTES defaults to 1 MiB per captured stdout/stderr stream. Client timeouts are clamped to the server maximum; omitted timeouts receive the default. Processes that time out or exceed the output cap are terminated and reaped before an explicit timeout/output-cap error is returned.
7. Audit Logging: Each invocation emits structured audit metadata for success, rejection, timeout, output cap, and process error outcomes. Secret-like argv values are redacted; stdout/stderr content is not logged.

Security-related environment variables

Variable	Default	Description
ALLOW_COMMANDS / ALLOWED_COMMANDS	empty	Comma-separated command names to allow
ALLOW_PATTERNS	empty	Comma-separated regex patterns matched with fullmatch() against command names
MCP_SHELL_DEFAULT_TIMEOUT_SECONDS	30	Timeout used when the client omits timeout
MCP_SHELL_MAX_TIMEOUT_SECONDS	300	Maximum effective timeout accepted from clients
MCP_SHELL_OUTPUT_LIMIT_BYTES	1048576	Maximum captured stdout/stderr bytes per process
MCP_SHELL_CHILD_ENV_ALLOWLIST	empty	Comma-separated parent or per-call environment variables allowed in children
MCP_SHELL_SAFE_PATH	/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin	PATH supplied to children and used for POSIX command lookup unless PATH is in MCP_SHELL_CHILD_ENV_ALLOWLIST; blank values use the default

Development

Setting up Development Environment

1. Clone the repository

git clone https://github.com/yourusername/mcp-shell-server.git
cd mcp-shell-server

2. Install dependencies including test requirements

pip install -e ".[test]"

Running Tests

pytest

API Reference

Request Arguments

Field	Type	Required	Description
command	string[]	Yes	Command and its arguments as array elements
stdin	string	No	Input to be passed to the command
directory	string	No	Working directory; omitted uses the server process CWD, and relative paths resolve from that server process CWD
timeout	integer	No	Maximum execution time in seconds

Response Fields

Field	Type	Description
stdout	string	Standard output from the command
stderr	string	Standard error output from the command
status	integer	Exit status code
execution_time	float	Time taken to execute (in seconds)
error	string	Error message (only present if failed)

Requirements

• Python 3.11 or higher
• mcp>=1.1.0

License

MIT License - See LICENSE file for details