pydantic-ai-filesystem-sandbox

Filesystem sandbox toolset for PydanticAI agents with LLM-friendly errors.

Why This Package?

When building LLM agents that interact with the filesystem, you need:

1. Sandboxing - Restrict which directories the agent can access
2. Read/Write Control - Fine-grained permissions per path
3. LLM-Friendly Errors - Error messages that help the LLM correct its behavior
4. Approval Integration - Works with human-in-the-loop approval flows

Architecture

• Sandbox - Security boundary for permission checking and path resolution
• FileSystemToolset - File I/O tools that use Sandbox
• ApprovalToolset - Optional wrapper for human-in-the-loop approval

Installation

pip install pydantic-ai-filesystem-sandbox

Quick Start

from pydantic_ai import Agent
from pydantic_ai_filesystem_sandbox import (
    FileSystemToolset,
    Sandbox,
    SandboxConfig,
    Mount,
)

# Configure sandbox with Docker-style mounts
config = SandboxConfig(mounts=[
    Mount(host_path="./data/input", mount_point="/input", mode="ro"),   # Read-only
    Mount(host_path="./data/output", mount_point="/output", mode="rw"), # Read-write
])

# Create the sandbox (security boundary)
sandbox = Sandbox(config)

# Create the toolset (file I/O tools)
toolset = FileSystemToolset(sandbox)

# Use with PydanticAI agent
agent = Agent("openai:gpt-4", toolsets=[toolset])

The agent can now access files using virtual paths like /input/data.txt and /output/results.json.

Simple Usage

For simple cases, use the factory method:

from pydantic_ai_filesystem_sandbox import FileSystemToolset

# Single directory with read-write access
toolset = FileSystemToolset.create_default("./data", mode="rw")

agent = Agent("openai:gpt-4", toolsets=[toolset])

Configuration

See API Reference for complete details.

Mount Options

| Option | Type | Default | Description | |--------|------|---------|-------------| | host_path | Path | required | Host directory to mount | | mount_point | str | required | Virtual path (e.g., "/docs", "/data") | | mode | "ro" | "rw" | "ro" | Access mode | | suffixes | list[str] | None | None | Allowed file extensions (None = all) | | max_file_bytes | int | None | None | Maximum file size limit | | write_approval | bool | True | Require approval for writes | | read_approval | bool | False | Require approval for reads |

Example Configuration

from pydantic_ai_filesystem_sandbox import Sandbox, SandboxConfig, Mount

config = SandboxConfig(mounts=[
    # Read-only input - any file type
    Mount(
        host_path="./data/input",
        mount_point="/input",
        mode="ro",
    ),
    # Read-write output - only markdown and text
    Mount(
        host_path="./data/output",
        mount_point="/output",
        mode="rw",
        suffixes=[".md", ".txt"],
        max_file_bytes=1_000_000,  # 1MB limit
    ),
    # Config files - read-only, requires approval
    Mount(
        host_path="./config",
        mount_point="/config",
        mode="ro",
        read_approval=True,
    ),
])

Single Root Mount

For projects where you want the entire directory as virtual /:

from pydantic_ai_filesystem_sandbox import Sandbox, SandboxConfig, Mount

config = SandboxConfig(mounts=[
    Mount(host_path=".", mount_point="/", mode="rw"),
])
sandbox = Sandbox(config)

sandbox.resolve("/src/main.py")  # -> <cwd>/src/main.py

Deriving child sandboxes

Use Sandbox.derive() to create a restricted child sandbox. By default the child has no access unless you explicitly allow paths.

parent = Sandbox(config)

# Empty child (secure by default)
child = parent.derive()

# Allow read-only access to a subtree
reader = parent.derive(allow_read="/output/reports")

# Allow read/write access to a subtree
writer = parent.derive(allow_write="/output/reports")

Available Tools

The toolset provides seven tools to the agent:

read_file

Read a text file from the sandbox.

Note: read_file currently reads the entire file into memory (even when using max_chars); use max_file_bytes to bound file size.

Path format: '/mount/path' (e.g., '/docs/file.txt')
Parameters:
  - path: str (required)
  - max_chars: int (default: 20,000)
  - offset: int (default: 0)

write_file

Write a text file to the sandbox (requires mode="rw"). Parent directories are created automatically.

Path format: '/mount/path' (e.g., '/output/file.txt')
Parameters:
  - path: str (required)
  - content: str (required)

edit_file

Edit a file by replacing exact text (requires mode="rw").

Note: edit_file reads the entire file into memory; use max_file_bytes to bound file size.

Path format: '/mount/path'
Parameters:
  - path: str (required)
  - old_text: str (required) - must match exactly and be unique
  - new_text: str (required)

delete_file

Delete a file from the sandbox (requires mode="rw").

Path format: '/mount/path'
Parameters:
  - path: str (required)

move_file

Move or rename a file within the sandbox (requires mode="rw" for both source and destination). Parent directories are created automatically.

Path format: '/mount/path'
Parameters:
  - source: str (required)
  - destination: str (required)

copy_file

Copy a file within the sandbox. Source can be read-only, destination requires mode="rw". Parent directories are created automatically.

Path format: '/mount/path'
Parameters:
  - source: str (required)
  - destination: str (required)

list_files

List files matching a glob pattern.

Parameters:
  - path: str (default: "/" for all mounts)
  - pattern: str (default: "**/*")

LLM-Friendly Errors

All errors include guidance on what IS allowed:

# PathNotInSandboxError
"Cannot access '/secret/file.txt': path is outside sandbox.
Readable paths: /input, /output"

# PathNotWritableError
"Cannot write to '/input/file.txt': path is read-only.
Writable paths: /output"

# SuffixNotAllowedError
"Cannot access '/output/data.json': suffix '.json' not allowed.
Allowed suffixes: .md, .txt"

# FileTooLargeError
"Cannot read '/output/huge.txt': file too large (5,000,000 bytes).
Maximum allowed: 1,000,000 bytes"

# EditError
"Cannot edit '/output/file.txt': text not found in file.
Searched for: 'old text...'"

Approval Integration

Works with pydantic-ai-blocking-approval for human-in-the-loop:

from pydantic_ai_filesystem_sandbox import (
    ApprovableFileSystemToolset, Sandbox, SandboxConfig, Mount
)
from pydantic_ai_blocking_approval import ApprovalToolset, ApprovalController

# Create sandbox and toolset
config = SandboxConfig(mounts=[
    Mount(host_path="./output", mount_point="/output", mode="rw", write_approval=True),
])
sandbox = Sandbox(config)
toolset = ApprovableFileSystemToolset(sandbox)

# Wrap with approval
controller = ApprovalController(mode="interactive")
approved_toolset = ApprovalToolset(
    inner=toolset,
    approval_callback=controller.approval_callback,
    memory=controller.memory,
)

agent = Agent(..., toolsets=[approved_toolset])

ApprovableFileSystemToolset extends FileSystemToolset with needs_approval() (returns ApprovalResult) and get_approval_description() for the approval UI.

Using the Sandbox Directly

The Sandbox class can be used independently for permission checking:

sandbox = Sandbox(config)

# Check permissions
if sandbox.can_write("/output/file.txt"):
    resolved = sandbox.resolve("/output/file.txt")
    # ... perform operation

# Query boundaries
print(sandbox.readable_roots)  # ["/input", "/output"]
print(sandbox.writable_roots)  # ["/output"]

# Check approval requirements
sandbox.needs_write_approval("/output/file.txt")  # True/False

ReadResult

The read_file tool returns a ReadResult object:

class ReadResult(BaseModel):
    content: str        # The file content read
    truncated: bool     # True if more content exists
    total_chars: int    # Total file size in characters
    offset: int         # Starting position used
    chars_read: int     # Characters actually returned

This allows agents to handle large files by reading in chunks:

# First read
result = toolset.read("/input/large.txt", max_chars=10000)
if result.truncated:
    # Continue reading
    result2 = toolset.read("/input/large.txt", max_chars=10000, offset=10000)

API Reference

See docs/api.md for full API documentation.

License

MIT