strands-shell 0.1.0

A virtual shell for AI agents

Strands Shell

Give your agent a shell without giving it the keys to your machine.

Documentation ◆ MCP Server ◆ Python ◆ Node.js

---

Agents run shell commands in tight loops: Install deps, run tests, grep for errors, iterate. Those loops need speed and isolation.

Strands Shell is a Bourne-compatible shell that runs in-process. grep, sed, jq, curl, find, 50+ commands and it does this without fork, exec, syscalls, or cold starts. You declare what the agent can reach (files, URLs, credentials) and everything else doesn't exist to the agent.

	Docker	Cloud sandbox	Strands Shell
Cold start	~200ms	~1s (network)	<1ms
Isolation	Container namespace	MicroVM	In-process VFS
Network	iptables / sidecar	Platform policy	URL allowlist + SSRF guard
Secrets	Env vars (agent can read them)	Platform-specific	Injected per-request, agent never sees them
Setup	Docker daemon	API key + network	pip install strands-shell
Platforms	Linux	Cloud-only	macOS, Linux, WASM

Quick Start

MCP (works with any agent framework)

Drop this into your MCP client config:

{
  "mcpServers": {
    "shell": {
      "command": "uvx",
      "args": ["strands-shell", "--mcp"]
    }
  }
}

That's it and your agent gets shell, read_file, write_file, list_dir. All mediated through the Kernel.

Python

pip install strands-shell

import strands_shell

shell = strands_shell.Shell(
    binds=[strands_shell.Bind("/my/project", "/workspace", mode="copy")],
    credentials=[strands_shell.Cred("https://api.example.com/", env_var="API_TOKEN")],
    allowed_urls=["https://api.example.com/"],
)

out = shell.run("grep -rn TODO /workspace")
print(out.stdout)

Node.js

npm install @strands-agents/shell

import { Shell } from '@strands-agents/shell'

const shell = await Shell.create({
  binds: [{ source: '/my/project', destination: '/workspace', mode: 'copy' }],
})
const out = await shell.run('grep -rn TODO /workspace')
console.log(out.stdout)

How It Works

flowchart TB
    agent["Your agent code\n(Strands, LangGraph, Pydantic AI, etc)"]
    agent -->|"MCP / Python / Node.js"| shell

    subgraph shell ["Strands Shell"]
        direction TB
        subgraph kernel ["Kernel (mediation boundary)"]
            vfs["VFS: isolated filesystem"]
            net["Network: SSRF guard + allowlist"]
            creds["Credentials: injected per-URL"]
            limits["Limits: timeout, output, fds"]
        end
        engine["Shell engine: parser, 25 builtins, 33 commands, Lua 5.4"]
    end

Written in Rust, with native bindings for Python (PyO3) and Node.js (napi-rs). State persists across run() calls (env vars, working directory, functions). The filesystem is shared.

Configuration

shell = strands_shell.Shell(
    binds=[
        strands_shell.Bind("/host/project", "/workspace", mode="copy"),
        strands_shell.Bind("/tmp/output", "/output", mode="direct"),
    ],
    credentials=[
        strands_shell.Cred("https://api.example.com/", env_var="API_TOKEN"),
    ],
    allowed_urls=["https://api.example.com/", "https://pypi.org/"],
    timeout=30.0,
    env={"PROJECT": "demo"},
    limits=strands_shell.Limits(
        max_output=1 << 20,
        max_file_size=10 << 20,
    ),
)

⚠️ mode: "direct" mounts are live. The agent can read and modify host files in real time. Use only for designated output directories. Never direct-bind directories containing secrets, credentials, or configuration you don't want the agent to modify.

TOML

You can load all of this from a config file instead:

[[bind]]
mode = "copy"
source = "/host/project"
destination = "/workspace"

[[cred]]
url = "https://api.openai.com/v1/"
methods = ["POST"]
kind = "bearer"
api_key_env = "OPENAI_API_KEY"

[[mcp]]
name = "my-tools"
command = "/path/to/mcp-server"
args = ["--stdio"]

MCP Server

The built-in MCP server exposes the shell over JSON-RPC on stdio, working with anything that speaks MCP.

uvx strands-shell --mcp                          # bare in-memory sandbox
uvx strands-shell --config sandbox.toml --mcp    # with mounts + credentials

If you declare [[mcp]] servers in your TOML config, they show up as Lua modules inside the shell. Call require("my_tools") and you get a table of the server's tools.

Security Model

Strands Shell is a mediation layer, not a security sandbox. It enforces what the agent should access via Kernel-mediated deny-by-default. It does NOT protect against: memory-safety exploits in the shell engine itself, timing side-channels, or an attacker who controls the host process. For multi-tenant or adversarial workloads, run each Shell instance inside a container or microVM.

The Kernel mediates everything; it runs in the same process as your code, not in a VM. If your threat model is "untrusted tenant running arbitrary code," put Strands Shell inside a container too. For "my agent shouldn't access things I haven't explicitly allowed," the Kernel handles it.

Default-deny. You allowlist what the agent can reach:

• Files: only bound paths exist, everything else is hidden.
• Network: curl blocks private ranges (RFC1918, link-local, loopback, IMDS) by default while letting public URLs pass through. Use allowed_urls to permit specific internal hosts.
• Secrets: the Kernel injects credentials per-URL at request time, ensuring the agent never holds them. The Kernel never re-injects on redirects, even back to the same host.
• Syscalls: there are none; no fork, no exec because the shell is pure userspace.

If you bypass any of these, report it. See SECURITY.md.

Limits (best-effort): timeouts, output caps, fd limits, inode limits. These catch runaway agents but won't stop someone actively trying to break out. OS-level isolation for that.

Multi-tenant: a Shell instance is single-owner. If you're serving multiple agents, create one Shell per session. Construction is cheap (no containers, no VMs, just an in-memory VFS), so spinning up per-request is the intended pattern.

Secure Defaults

Out of the box, the shell is an empty sandbox — no files, no network, no credentials. When you grant access, follow least privilege:

• Prefer mode: "copy" over mode: "direct" for source code. Copy-on-create isolates the agent from your live files. Use direct only for output directories where the agent needs to persist results.
• Scope binds narrowly. Bind /my/project/src rather than /my/project or /. The agent doesn't need your .git/, .env, or node_modules/.
• Allowlist URLs explicitly. Don't use allowed_urls: ["https://"] — this disables SSRF protection entirely. List the specific API endpoints the agent needs.
• Set timeouts. The default has no per-command timeout. Set timeout to bound runaway commands (30s is reasonable for most agent loops).
• Use limits. Set max_output to prevent agents from filling memory with unbounded command output (1MB is a good default).

Commands

25 builtins, 33 commands, and a Bourne-compatible shell with pipes, loops, functions, and subshells.

The commands agents use constantly: grep, find, cat, head, tail, jq for reading and searching. sed, sort, tr, cut for transforming output. cp, mv, rm, mkdir for managing files. curl for HTTP (SSRF-guarded, credentials auto-injected). lua for scripting when shell gets awkward.

COMMANDS.md has the full inventory with implementation status, supported flags, and known gaps vs GNU coreutils.

File Operations API

Read and write files without going through a shell command:

shell.write_file("/workspace/note.txt", b"hello")
data = shell.read_file("/workspace/note.txt")
entries = shell.list_files("/workspace")
shell.remove_file("/workspace/note.txt")

Contributing

See CONTRIBUTING.md. Bug reports and design questions are just as useful as PRs.

Community

Discord if you want to talk about it.

License

Apache-2.0

Security

If you find a security issue, report it privately instead of opening a public issue. Bypasses of filesystem mediation, SSRF protection, or credential injection qualify. See SECURITY.md.