Bash tool

MessagesTools

Bash tool

This feature is eligible for Zero Data Retention (ZDR). When your organization has a ZDR arrangement, data sent through this feature is not stored after the API response is returned.

The bash tool enables Claude to execute shell commands in a persistent bash session, allowing system operations, script execution, and command-line automation. Shell access is a foundational agent capability. On Terminal-Bench 2.0, a benchmark that evaluates real-world terminal tasks using shell-only validation, Claude shows strong performance gains with access to a persistent bash session.

Overview

The bash tool provides Claude with:

Persistent bash session that maintains state
Ability to run any shell command
Access to environment variables and working directory
Command chaining and scripting capabilities

For model support, see the Tool reference.

Use cases

Development workflows: Run build commands, tests, and development tools
System automation: Execute scripts, manage files, automate tasks
Data processing: Process files, run analysis scripts, manage datasets
Environment setup: Install packages, configure environments

Quick start

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    tools=[{"type": "bash_20250124", "name": "bash"}],
    messages=[
        {"role": "user", "content": "List all Python files in the current directory."}
    ],
)

print(response)

How it works

The bash tool maintains a persistent session:

Claude determines what command to run
You execute the command in a bash shell
Return the output (stdout and stderr) to Claude
Session state persists between commands (environment variables, working directory)

Parameters

Parameter	Required	Description
`command`	Yes*	The bash command to run
`restart`	No	Set to `true` to restart the bash session

*Required unless using restart

Example: Multi-step automation

Claude can chain commands to complete complex tasks:

User request:
"Install the requests library and create a simple Python script that
fetches a joke from an API, then run it."

Claude's tool uses:
1. Install package
   {"command": "pip install requests"}

2. Create script
   {"command": "cat > fetch_joke.py << 'EOF'\nimport requests\nresponse = requests.get('https://official-joke-api.appspot.com/random_joke')\njoke = response.json()\nprint(f\"Setup: {joke['setup']}\")\nprint(f\"Punchline: {joke['punchline']}\")\nEOF"}

3. Run script
   {"command": "python fetch_joke.py"}

The session maintains state between commands, so files created in step 2 are available in step 3.

Implement the bash tool

The bash tool is implemented as a schema-less tool. When using this tool, you don't need to provide an input schema as with other tools; the schema is built into Claude's model and can't be modified.

Set up a bash environment

Create a persistent bash session that Claude can interact with:

import subprocess
import threading
import queue


class BashSession:
    def __init__(self):
        self.process = subprocess.Popen(
            ["/bin/bash"],
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
            bufsize=0,
        )
        self.output_queue = queue.Queue()
        self.error_queue = queue.Queue()
        self._start_readers()

Handle command execution

Create a function to execute commands and capture output:

def execute_command(self, command):
    # Send command to bash
    self.process.stdin.write(command + "\n")
    self.process.stdin.flush()

    # Capture output with timeout
    output = self._read_output(timeout=10)
    return output

Process Claude's tool calls

Extract and execute commands from Claude's responses:

for content in response.content:
    if content.type == "tool_use" and content.name == "bash":
        if content.input.get("restart"):
            bash_session.restart()
            result = "Bash session restarted"
        else:
            command = content.input.get("command")
            result = bash_session.execute_command(command)

        # Return result to Claude
        tool_result = {
            "type": "tool_result",
            "tool_use_id": content.id,
            "content": result,
        }

Implement safety measures

Add validation and restrictions. Use an allowlist rather than a blocklist, since blocklists are easy to bypass. Reject shell operators so chained commands can't slip past the allowlist:

import shlex

ALLOWED_COMMANDS = {"ls", "cat", "echo", "pwd", "grep", "find", "wc", "head", "tail"}
SHELL_OPERATORS = {"&&", "||", "|", ";", "&", ">", "<", ">>"}


def validate_command(command):
    # Allow only commands from an explicit allowlist
    try:
        tokens = shlex.split(command)
    except ValueError:
        return False, "Could not parse command"

    if not tokens:
        return False, "Empty command"

    executable = tokens[0]
    if executable not in ALLOWED_COMMANDS:
        return False, f"Command '{executable}' is not in the allowlist"

    # Reject shell operators that would chain additional commands
    for token in tokens[1:]:
        if token in SHELL_OPERATORS or token.startswith(("$", "`")):
            return False, f"Shell operator '{token}' is not allowed"

    return True, None

This check is a first line of defense. For stronger isolation, run validated commands with shell=False and pass shlex.split(command) as the argument list, so the shell never interprets the string.

Handle errors

When implementing the bash tool, handle various error scenarios:

Follow implementation best practices

Security

The bash tool provides direct system access. Implement these essential safety measures:

Running in isolated environments (Docker/VM)
Implementing command filtering and allowlists
Setting resource limits (CPU, memory, disk)
Logging all executed commands

Key recommendations

Use ulimit to set resource constraints
Filter dangerous commands (sudo, rm -rf, etc.)
Run with minimal user permissions
Monitor and log all command execution

Pricing

The bash tool adds 245 input tokens to your API calls.

Additional tokens are consumed by:

Command outputs (stdout/stderr)
Error messages
Large file contents

See tool use pricing for complete pricing details.

Common patterns

Development workflows

Running tests: pytest && coverage report
Building projects: npm install && npm run build
Git operations: git status && git add . && git commit -m "message"

Git-based checkpointing

Git serves as a structured recovery mechanism in long-running agent workflows, not just a way to save changes:

Capture a baseline: Before any agent work begins, commit the current state. This is the known-good starting point.
Commit per feature: Each completed feature gets its own commit. These serve as rollback points if something goes wrong later.
Reconstruct state at session start: Read git log alongside a progress file to understand what has already been done and what comes next.
Revert on failure: If work goes sideways, git checkout reverts to the last good commit instead of trying to debug a broken state.

File operations

Processing data: wc -l *.csv && ls -lh *.csv
Searching files: find . -name "*.py" | xargs grep "pattern"
Creating backups: tar -czf backup.tar.gz ./data

System tasks

Checking resources: df -h && free -m
Process management: ps aux | grep python
Environment setup: export PATH=$PATH:/new/path && echo $PATH

Limitations

No interactive commands: Cannot handle vim, less, or password prompts
No GUI applications: Command-line only
Session scope: Bash session state is client-side. The API is stateless. Your application is responsible for maintaining the shell session between turns.
Output limits: Large outputs may be truncated
No streaming: Results returned after completion

Combining with other tools

The bash tool is most powerful when combined with the text editor and other tools.

If you're also using the code execution tool, Claude has access to two separate execution environments: your local bash session and Anthropic's sandboxed container. State is not shared between them. See Using code execution with other execution tools for guidance on prompting Claude to distinguish between environments.

Next steps

Tool use overview

Learn about tool use with Claude

Text editor tool

View and edit text files with Claude

Was this page helpful?

MessagesTools

Bash tool

This feature is eligible for Zero Data Retention (ZDR). When your organization has a ZDR arrangement, data sent through this feature is not stored after the API response is returned.

Overview

The bash tool provides Claude with:

Persistent bash session that maintains state
Ability to run any shell command
Access to environment variables and working directory
Command chaining and scripting capabilities

For model support, see the Tool reference.

Use cases

Development workflows: Run build commands, tests, and development tools
System automation: Execute scripts, manage files, automate tasks
Data processing: Process files, run analysis scripts, manage datasets
Environment setup: Install packages, configure environments

Quick start

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    tools=[{"type": "bash_20250124", "name": "bash"}],
    messages=[
        {"role": "user", "content": "List all Python files in the current directory."}
    ],
)

print(response)

How it works

The bash tool maintains a persistent session:

Claude determines what command to run
You execute the command in a bash shell
Return the output (stdout and stderr) to Claude
Session state persists between commands (environment variables, working directory)

Parameters

Parameter	Required	Description
`command`	Yes*	The bash command to run
`restart`	No	Set to `true` to restart the bash session

*Required unless using restart

Example: Multi-step automation

Claude can chain commands to complete complex tasks:

User request:
"Install the requests library and create a simple Python script that
fetches a joke from an API, then run it."

Claude's tool uses:
1. Install package
   {"command": "pip install requests"}

2. Create script
   {"command": "cat > fetch_joke.py << 'EOF'\nimport requests\nresponse = requests.get('https://official-joke-api.appspot.com/random_joke')\njoke = response.json()\nprint(f\"Setup: {joke['setup']}\")\nprint(f\"Punchline: {joke['punchline']}\")\nEOF"}

3. Run script
   {"command": "python fetch_joke.py"}

The session maintains state between commands, so files created in step 2 are available in step 3.

Implement the bash tool

The bash tool is implemented as a schema-less tool. When using this tool, you don't need to provide an input schema as with other tools; the schema is built into Claude's model and can't be modified.

Set up a bash environment

Create a persistent bash session that Claude can interact with:

import subprocess
import threading
import queue


class BashSession:
    def __init__(self):
        self.process = subprocess.Popen(
            ["/bin/bash"],
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
            bufsize=0,
        )
        self.output_queue = queue.Queue()
        self.error_queue = queue.Queue()
        self._start_readers()

Handle command execution

Create a function to execute commands and capture output:

def execute_command(self, command):
    # Send command to bash
    self.process.stdin.write(command + "\n")
    self.process.stdin.flush()

    # Capture output with timeout
    output = self._read_output(timeout=10)
    return output

Process Claude's tool calls

Extract and execute commands from Claude's responses:

for content in response.content:
    if content.type == "tool_use" and content.name == "bash":
        if content.input.get("restart"):
            bash_session.restart()
            result = "Bash session restarted"
        else:
            command = content.input.get("command")
            result = bash_session.execute_command(command)

        # Return result to Claude
        tool_result = {
            "type": "tool_result",
            "tool_use_id": content.id,
            "content": result,
        }

Implement safety measures

Add validation and restrictions. Use an allowlist rather than a blocklist, since blocklists are easy to bypass. Reject shell operators so chained commands can't slip past the allowlist:

import shlex

ALLOWED_COMMANDS = {"ls", "cat", "echo", "pwd", "grep", "find", "wc", "head", "tail"}
SHELL_OPERATORS = {"&&", "||", "|", ";", "&", ">", "<", ">>"}


def validate_command(command):
    # Allow only commands from an explicit allowlist
    try:
        tokens = shlex.split(command)
    except ValueError:
        return False, "Could not parse command"

    if not tokens:
        return False, "Empty command"

    executable = tokens[0]
    if executable not in ALLOWED_COMMANDS:
        return False, f"Command '{executable}' is not in the allowlist"

    # Reject shell operators that would chain additional commands
    for token in tokens[1:]:
        if token in SHELL_OPERATORS or token.startswith(("$", "`")):
            return False, f"Shell operator '{token}' is not allowed"

    return True, None

Handle errors

When implementing the bash tool, handle various error scenarios:

Follow implementation best practices

Security

The bash tool provides direct system access. Implement these essential safety measures:

Running in isolated environments (Docker/VM)
Implementing command filtering and allowlists
Setting resource limits (CPU, memory, disk)
Logging all executed commands

Key recommendations

Use ulimit to set resource constraints
Filter dangerous commands (sudo, rm -rf, etc.)
Run with minimal user permissions
Monitor and log all command execution

Pricing

The bash tool adds 245 input tokens to your API calls.

Additional tokens are consumed by:

Command outputs (stdout/stderr)
Error messages
Large file contents

See tool use pricing for complete pricing details.

Common patterns

Development workflows

Running tests: pytest && coverage report
Building projects: npm install && npm run build
Git operations: git status && git add . && git commit -m "message"

Git-based checkpointing

Git serves as a structured recovery mechanism in long-running agent workflows, not just a way to save changes:

Capture a baseline: Before any agent work begins, commit the current state. This is the known-good starting point.
Commit per feature: Each completed feature gets its own commit. These serve as rollback points if something goes wrong later.
Reconstruct state at session start: Read git log alongside a progress file to understand what has already been done and what comes next.
Revert on failure: If work goes sideways, git checkout reverts to the last good commit instead of trying to debug a broken state.

File operations

Processing data: wc -l *.csv && ls -lh *.csv
Searching files: find . -name "*.py" | xargs grep "pattern"
Creating backups: tar -czf backup.tar.gz ./data

System tasks

Checking resources: df -h && free -m
Process management: ps aux | grep python
Environment setup: export PATH=$PATH:/new/path && echo $PATH

Limitations

No interactive commands: Cannot handle vim, less, or password prompts
No GUI applications: Command-line only
Session scope: Bash session state is client-side. The API is stateless. Your application is responsible for maintaining the shell session between turns.
Output limits: Large outputs may be truncated
No streaming: Results returned after completion

Combining with other tools

The bash tool is most powerful when combined with the text editor and other tools.

Next steps

Tool use overview

Learn about tool use with Claude

Text editor tool

View and edit text files with Claude

Was this page helpful?

Overview

Use cases

Quick start

How it works

Parameters

Example usage

Example: Multi-step automation

Implement the bash tool

Handle errors

Command execution timeout

Command not found

Permission denied

Follow implementation best practices

Use command timeouts

Maintain session state

Handle large outputs

Log all commands

Sanitize outputs

Security

Key recommendations

Pricing

Common patterns

Development workflows

Git-based checkpointing

File operations

System tasks

Limitations

Combining with other tools

Next steps

Overview

Use cases

Quick start

How it works

Parameters

Example usage

Example: Multi-step automation

Implement the bash tool

Handle errors

Command execution timeout

Command not found

Permission denied

Follow implementation best practices

Use command timeouts

Maintain session state

Handle large outputs

Log all commands

Sanitize outputs

Security

Key recommendations

Pricing

Common patterns

Development workflows

Git-based checkpointing

File operations

System tasks

Limitations

Combining with other tools

Next steps