rshell - A Restricted Shell for AI Agents

A restricted shell interpreter for Go. Designed for AI agents that need to run shell commands safely.

Install

go get github.com/DataDog/rshell

Quick Start

package main

import (
	"context"
	"os"
	"strings"
	"time"

	"github.com/DataDog/rshell/interp"
	"mvdan.cc/sh/v3/syntax"
)

func main() {
	script := `echo "hello from rshell"`

	prog, _ := syntax.NewParser().Parse(strings.NewReader(script), "")

	runner, _ := interp.New(
		interp.StdIO(nil, os.Stdout, os.Stderr),
		interp.AllowedCommands([]string{"rshell:echo"}),
		interp.MaxExecutionTime(5*time.Second),
	)
	defer runner.Close()

	runner.Run(context.Background(), prog)
}

CLI usage also supports a whole-run timeout:

rshell --allow-all-commands --timeout 5s -c 'echo "hello from rshell"'

Security Model

CLI scope: The rshell CLI is a development, debugging, and local validation harness for the interpreter. It is not intended to be exposed as a production execution boundary or as an interface for untrusted users. Production integrations should embed the Go API and set AllowedCommands, AllowedPaths, environment, timeout, and mode explicitly. Security review should treat the interpreter and embedding configuration as the security boundary, not the developer CLI.

Every access path is default-deny:

Resource	Default	Opt-in
Command execution	All commands blocked (exit code 127)	AllowedCommands with namespaced command list (e.g. rshell:cat)
Systemd services	All services and actions blocked	AllowedSystemServices with exact service/action grants
External commands	Blocked (exit code 127)	Provide an ExecHandler
Filesystem access	Blocked	Configure AllowedPaths with `PATH[:ro
Environment variables	Empty (no host env inherited)	Pass variables via the Env option
Output redirections	Only /dev/null allowed in read-only mode (exit code 2 for other targets); file-target redirects enabled in remediation mode, gated by AllowedPaths	>/dev/null, 2>/dev/null, &>/dev/null, 2>&1; in remediation mode: >FILE, >>FILE, 2>FILE, &>FILE, &>>FILE

AllowedCommands restricts which commands (builtins or external) the interpreter may execute. Commands must be specified with the rshell: namespace prefix (e.g. rshell:cat, rshell:echo). If not set, no commands are allowed.

AllowedSystemServices is the single capability policy shared by systemd-aware builtins. Grants pair one exact service with generic actions (read, clean, reload, or restart), using SERVICE:ACTION[+ACTION...] syntax. Grants without actions are ignored; invalid services and unsupported actions are skipped with warnings. Service names are matched exactly without adding suffixes, resolving aliases, changing case, or otherwise normalizing them: mysql and mysql.service are different services. Empty service names, service names containing :, whitespace, path-like names, and glob patterns are also skipped with warnings. The policy defaults to denying every operation and remains enforced when all commands are allowed. read is available in read-only mode; mutating actions require remediation mode.

interp.AllowedSystemServices([]interp.SystemdControlGrant{
	{
		Service: "mysql.service",
		Actions: []interp.SystemServiceAction{
			interp.SystemServiceRestart,
			interp.SystemServiceReload,
			interp.SystemServiceRead,
		},
	},
	{
		Service: "systemd-journald.service",
		Actions: []interp.SystemServiceAction{interp.SystemServiceRead, interp.SystemServiceClean},
	},
})

The development CLI accepts equivalent grants through --allowed-services mysql.service:restart+reload+read,systemd-journald.service:read+clean. Service selectors are always exact service names; : separates a selector from its actions and is not allowed inside service names. The restricted journalctl builtin is available as described below; systemctl is not yet implemented.

SystemdTargetConfig selects which Linux host journal-aware builtins address. With no option, standard local paths are used. Container integrations can instead provide JournalDirs, MachineIDPath, and JournalControlSocket as direct absolute paths visible to the rshell process. Once any explicit field is supplied, omitted fields stay unavailable and never fall back to local endpoints. MachineIDPath is mandatory for every explicit target. The development CLI exposes the equivalent --systemd-journal-dirs, --systemd-machine-id-path, and --systemd-journal-socket flags.

The target paths intentionally bypass AllowedPaths: they are trusted runner configuration and cannot be supplied by shell scripts. The embedding application is responsible for mounting every supplied path from the same host. Every selected journal file header is checked against the configured machine ID, but journald's Rotate Varlink method does not return a machine ID that can independently attest its socket.

Operation	Required target access
Journal query, disk usage, or vacuum	/etc/machine-id and one or both of /var/log/journal, /run/log/journal
Journal rotation	/etc/machine-id and /run/systemd/journal/io.systemd.journal

Explicit targets may map those host locations to arbitrary absolute container paths. A command fails closed when one of its required paths was omitted.

Restricted journalctl

journalctl provides bounded investigation and disk-cleanup operations without executing the host binary or accepting user-selected files, directories, journal matches, cursors, machines, or namespaces.

Operation	Supported flags	Required systemd grant
Exact service logs	-u SERVICE (repeatable), -b, -n COUNT, --since TIME, -o short|cat	Exact SERVICE:read grant for every service
Current-boot kernel logs	-k, -n COUNT, --since TIME, -o short|cat	systemd-journald.service:read
Allocated journal usage	--disk-usage	systemd-journald.service:read
Synchronous active-file rotation	--rotate	systemd-journald.service:clean plus remediation mode
Archived-file cleanup	--vacuum-size SIZE, --vacuum-time AGE, --dry-run	systemd-journald.service:clean plus remediation mode

Log queries default to 100 entries and are capped at 1,000 entries and 32 exact unit scopes per invocation. --since accepts RFC 3339, local YYYY-MM-DD HH:MM:SS, or a non-negative Go lookback duration such as 15m. -b means the newest boot present in the selected target journal, so it remains correct for a mounted host. Bare journal reads, globbed units, arbitrary field matches, follow mode, raw structured output, alternate sources, and arbitrary boot selection are unavailable.

Selected journal fields are capped at 64 KiB. short escapes embedded newlines, carriage returns, tabs, non-graphic Unicode code points, and invalid UTF-8 bytes before writing output; kernel entries without an identifier are labeled kernel, matching host journalctl -k. Within that bound, cat matches host journalctl -o cat by writing each MESSAGE value unchanged and appending one newline; callers must treat that raw output as untrusted log content.

Log reading uses a bounded pure-Go journal-file parser and does not execute host journalctl or require cgo or libsystemd. It supports regular and compact journal layouts, legacy and keyed DATA hashes, and XZ, LZ4, and Zstandard-compressed DATA objects. Unknown incompatible format features fail with a clear error.

Unit reads include journalctl's trusted direct-unit, manager, root object-unit, root coredump, and slice sources. The reader opens at most 128 regular, non-symlink journal files, verifies the machine ID plus each file's identity, size, modification time, and parsed header before and after each query, and retries once if concurrent rotation or an in-progress write makes the first snapshot inconsistent. Rotation requires Linux, the configured journald Varlink socket, and procfs descriptor links at /proc/self/fd. The backend pins the socket inode before connecting through its descriptor, so a concurrent path replacement cannot redirect the request; rotation fails closed when descriptor links are unavailable. Disk usage and vacuuming use the mounted journal files directly on Linux or macOS. Used alone, --vacuum-time removes every eligible archive at least that old. --vacuum-size requires --vacuum-time; when combined, the age becomes a minimum deletion age and cleanup stops once the size target is reached, leaving newer archives intact even when usage remains above the target. --rotate may be combined with vacuum flags, but newly rotated archives remain protected by the time floor. --dry-run cannot be combined with --rotate and still requires the clean grant and remediation mode.

Vacuum thresholds are provided by the journalctl command itself. The backend removes only strict systemd archived-journal and .journal~ corruption filenames, never active files, symlinks, hardlinks, or malformed names, and it revalidates each file immediately before deletion.

AllowedPaths restricts all file operations to specified directories using Go's os.Root API for reads and openat-based write handling for writes.

• Sandbox mechanism: Reads go through os.Root; writes are checked against the most-specific path mode and, on Unix, opened with a no-symlink openat walk. Files outside the allowlist cannot be opened, created, truncated, or appended to.
• Permission suffix: Path entries may end with :ro or :rw representing read-only and read-write modes, respectively; entries without a suffix default to read-only, and the suffix is stripped before path validation. In remediation mode, write operations are accepted only inside the most-specific matching :rw root.
• Symlink policy: A symlink pointing outside its os.Root is followed for reads but never for writes; on Unix, symlink components in write targets are rejected with symlinks are not supported as write targets rather than followed, eliminating the TOCTOU window where a malicious link target could be swapped between resolution and open.
• Output redirections by mode:
  • Read-only mode (default): file-target output redirections (>, >>, 2>, &>, &>>) are rejected at parse time (exit 2).
  • Remediation mode: those redirections open through the same sandbox — writes inside :rw allowlist roots succeed; targets outside the allowlist or inside read-only roots fail with permission denied (exit 1).
• Special targets: The literal target /dev/null is always short-circuited to a discarded sink without going through the sandbox. <> (read-write open) is blocked in all modes.
• Diagnostic messages: Configured directories that cannot be opened and invalid system service names are skipped with a warning flushed once to the runner's stderr at construction time. Silence them or redirect them programmatically with WarningsWriter(io.Writer) or inspect them with Runner.Warnings().

Note: The ss, ip route, and df builtins bypass AllowedPaths for their kernel-state reads. ss and ip route open /proc/net/* paths directly; df reads /proc/self/mountinfo (Linux) or calls getfsstat(2) (macOS), then issues unix.Statfs(2) against every kernel-reported mount point. These paths are hardcoded — never derived from user input — and Statfs returns metadata only (block / inode counts, filesystem type, block size). There is no sandbox-escape risk, but operators cannot use AllowedPaths to block ss from enumerating local sockets, ip route from reading the routing table, or df from reporting mount-table capacity — these reads succeed regardless of the configured path policy.

ProcPath (Linux-only) overrides the proc filesystem root used by the ps builtin (default /proc). This is a privileged option set at runner construction time by trusted caller code — scripts cannot influence it. Access to the proc path is intentionally not subject to AllowedPaths restrictions. To avoid leaking secrets passed as CLI arguments, ps does not read /proc/<pid>/cmdline; the CMD column reports only the process comm/executable name.

RemediationMode opts the runner into host-remediation mode, enabling file-target output redirections (>, >>, 2>, &>, &>>) within :rw entries in the configured AllowedPaths and enabling remediation-only builtins such as truncate and logrotate. Targets outside the allowlist or inside read-only roots are rejected with permission denied (exit 1); symlinked write targets are rejected with symlinks are not supported as write targets; /dev/null is always accepted. <> (read-write open) remains blocked in all modes. CLI flag: --mode remediation (default: --mode read-only). The logrotate builtin is an rshell-safe log truncation helper, not a full logrotate(8) replacement.

Shell Features

Inside rshell, run help to list supported feature categories, a concise unsupported-feature summary, enabled commands, and the configured AllowedPaths sandbox roots grouped by read-only and read-write access (or a notice when none are configured). Use help <feature|command> for details about a specific rshell feature or command.

See SHELL_FEATURES.md for the complete list of supported and blocked features.

Platform Support

Linux, macOS, and Windows.

Publishing Changes

After merging changes to main create a release by:

1. Navigate to the Releases page

2. Click "Draft a new release"

3. You can "Select a tag" using the dropdown or "Create a new tag"

   When creating a new tag, make sure to include the v prefix. For example, if the last release was v0.1.29, your release should be v0.1.30.

4. The release title should be the same as the version tag

5. Use "Generate release notes" to fill in the release description

6. Click "Publish release"

   This will create a git tag that can now be referenced in other repos. This will trigger go-releaser that will add installable artifacts to the release.

License

Apache License 2.0