bashkit 0.1.15

A sandboxed bash interpreter for AI agents

Bashkit

A sandboxed bash interpreter for AI agents.

from bashkit import BashTool

tool = BashTool()
result = tool.execute_sync("echo 'Hello, World!'")
print(result.stdout)  # Hello, World!

Features

• Sandboxed execution — all commands run in-process with a virtual filesystem, no containers needed
• 156 built-in commands — echo, cat, grep, sed, awk, jq, curl, find, and more
• Full bash syntax — variables, pipelines, redirects, loops, functions, arrays
• Resource limits — protect against infinite loops and runaway scripts
• Framework integrations — LangChain, PydanticAI, and Deep Agents

Installation

pip install bashkit

# With framework support
pip install 'bashkit[langchain]'
pip install 'bashkit[pydantic-ai]'

Usage

Async

import asyncio
from bashkit import Bash

async def main():
    bash = Bash()

    # Simple command
    result = await bash.execute("echo 'Hello, World!'")
    print(result.stdout)  # Hello, World!

    # Pipeline
    result = await bash.execute("echo -e 'banana\\napple\\ncherry' | sort")
    print(result.stdout)  # apple\nbanana\ncherry

    # Virtual filesystem persists between calls
    await bash.execute("echo 'data' > /tmp/file.txt")
    result = await bash.execute("cat /tmp/file.txt")
    print(result.stdout)  # data

asyncio.run(main())

Sync

from bashkit import BashTool

tool = BashTool()
result = tool.execute_sync("echo 'Hello!'")
print(result.stdout)

Configuration

bash = Bash(
    username="agent",           # Custom username (whoami)
    hostname="sandbox",         # Custom hostname
    max_commands=1000,          # Limit total commands
    max_loop_iterations=10000,  # Limit loop iterations
)

Direct Filesystem Access

from bashkit import Bash

bash = Bash()
fs = bash.fs()
fs.mkdir("/data", recursive=True)
fs.write_file("/data/blob.bin", b"\x00\x01hello")
assert fs.read_file("/data/blob.bin") == b"\x00\x01hello"

# Additional filesystem operations
fs.append_file("/data/blob.bin", b" world")
fs.copy("/data/blob.bin", "/data/backup.bin")
fs.rename("/data/backup.bin", "/data/copy.bin")
info = fs.stat("/data/blob.bin")       # dict with size, type, etc.
entries = fs.read_dir("/data")         # detailed directory listing
fs.symlink("/data/link", "/data/blob.bin")
fs.chmod("/data/blob.bin", 0o644)

Live Mounts

from bashkit import Bash, FileSystem

bash = Bash()
workspace = FileSystem.real("/path/to/workspace", readwrite=True)
bash.mount("/workspace", workspace)

bash.execute_sync("echo 'hello' > /workspace/demo.txt")
bash.unmount("/workspace")

Bash and BashTool also still support constructor-time mounts:

• mount_text=[("/path", "content")]
• mount_readonly_text=[("/path", "content")]
• mount_real_readonly=["/host/path"]
• mount_real_readonly_at=[("/host/path", "/vfs/path")]
• mount_real_readwrite=["/host/path"]
• mount_real_readwrite_at=[("/host/path", "/vfs/path")]

BashTool — Convenience Wrapper for AI Agents

BashTool is a convenience wrapper specifically designed for AI agents. It wraps Bash and adds contract metadata (description, Markdown help, system_prompt, JSON schemas) needed by tool-use protocols. Use this when integrating with LangChain, PydanticAI, or similar agent frameworks.

from bashkit import BashTool

tool = BashTool()
print(tool.input_schema())    # JSON schema for LLM tool-use
print(tool.description())     # Token-efficient tool description
print(tool.system_prompt())   # Token-efficient prompt
print(tool.help())            # Markdown help document

result = await tool.execute("echo 'Hello!'")

Scripted Tool Orchestration

Compose multiple tools into a single bash-scriptable interface:

from bashkit import ScriptedTool

tool = ScriptedTool("api")
tool.add_tool("greet", "Greet a user", callback=lambda p, s=None: f"hello {p.get('name', 'world')}")
result = tool.execute_sync("greet --name Alice")
print(result.stdout)  # hello Alice

LangChain

from bashkit.langchain import create_bash_tool

bash_tool = create_bash_tool()
# Use with any LangChain agent

PydanticAI

from bashkit.pydantic_ai import create_bash_tool

bash_tool = create_bash_tool()
# Use with any PydanticAI agent

Deep Agents

from bashkit.deepagents import BashkitBackend, BashkitMiddleware
# Use with Deep Agents framework

ScriptedTool — Multi-Tool Orchestration

Compose Python callbacks as bash builtins. An LLM writes a single bash script that pipes, loops, and branches across all registered tools.

from bashkit import ScriptedTool

def get_user(params, stdin=None):
    return '{"id": 1, "name": "Alice"}'

tool = ScriptedTool("api")
tool.add_tool("get_user", "Fetch user by ID",
    callback=get_user,
    schema={"type": "object", "properties": {"id": {"type": "integer"}}})

result = tool.execute_sync("get_user --id 1 | jq -r '.name'")
print(result.stdout)  # Alice

Features

• Sandboxed, in-process execution: All commands run in isolation with a virtual filesystem
• 150 built-in commands: echo, cat, grep, sed, awk, jq, curl, find, and more
• Full bash syntax: Variables, pipelines, redirects, loops, functions, arrays
• Resource limits: Protect against infinite loops and runaway scripts

API Reference

Bash

• execute(commands: str) -> ExecResult — execute commands asynchronously
• execute_sync(commands: str) -> ExecResult — execute commands synchronously
• execute_or_throw(commands: str) -> ExecResult — async, raises on non-zero exit
• execute_sync_or_throw(commands: str) -> ExecResult — sync, raises on non-zero exit
• reset() — reset interpreter state
• fs() -> FileSystem — direct filesystem access

BashTool

Convenience wrapper for AI agents. Inherits all execution methods from Bash, plus:

• description() -> str — token-efficient tool description
• help() -> str — Markdown help document
• system_prompt() -> str — token-efficient system prompt for LLM integration
• input_schema() -> str — JSON input schema
• output_schema() -> str — JSON output schema

ExecResult

• stdout: str — standard output
• stderr: str — standard error
• exit_code: int — exit code (0 = success)
• error: Optional[str] — error message if execution failed
• success: bool — True if exit_code == 0
• to_dict() -> dict — convert to dictionary

FileSystem

• mkdir(path, recursive=False) — create directory
• write_file(path, content) — write bytes to file
• read_file(path) -> bytes — read file contents
• append_file(path, content) — append bytes to file
• exists(path) -> bool — check if path exists
• remove(path, recursive=False) — delete file/directory
• stat(path) -> dict — file metadata (size, type, etc.)
• read_dir(path) -> list — detailed directory listing
• rename(src, dst) — rename file/directory
• copy(src, dst) — copy file
• symlink(link, target) — create symlink
• chmod(path, mode) — change file permissions
• read_link(path) -> str — read symlink target

ScriptedTool

• add_tool(name, description, callback, schema=None) — register a tool
• execute(script: str) -> ExecResult — execute script asynchronously
• execute_sync(script: str) -> ExecResult — execute script synchronously
• execute_or_throw(script: str) -> ExecResult — async, raises on non-zero exit
• execute_sync_or_throw(script: str) -> ExecResult — sync, raises on non-zero exit
• env(key: str, value: str) — set environment variable
• tool_count() -> int — number of registered tools

How it works

Bashkit is built on top of Bashkit core, a bash interpreter written in Rust. The Python package provides a native extension for fast, sandboxed execution without spawning subprocesses or containers.

Part of Everruns

Bashkit is part of the Everruns ecosystem — tools and runtimes for building reliable AI agents. See the bashkit monorepo for the Rust core, Node.js package (@everruns/bashkit), and more.

License

MIT