pydantic-ai-backend 0.2.13

File storage and sandbox backends for AI agents

File Storage & Sandbox Backends for Pydantic AI

Console Toolset, Docker Sandbox, and Permission System for AI Agents

Console Toolset — ls, read, write, edit, grep, execute  •  Docker Sandbox — isolated code execution  •  Permission System — fine-grained access control

---

File Storage & Sandbox Backends provides everything your Pydantic AI agent needs to work with files and execute code safely. Choose from in-memory, local filesystem, or Docker-isolated backends.

Full framework? Check out Pydantic Deep Agents — complete agent framework with planning, filesystem, subagents, and skills.

Use Cases

What You Want to Build	How This Library Helps
AI Coding Assistant	Console toolset with file ops + code execution
Multi-User Web App	Docker sandboxes with session isolation
Code Review Bot	Read-only backend with grep/glob search
Secure Execution	Permission system blocks dangerous operations
Testing/CI	In-memory StateBackend for fast, isolated tests

Installation

pip install pydantic-ai-backend

Or with uv:

uv add pydantic-ai-backend

Optional extras:

# Console toolset (requires pydantic-ai)
pip install pydantic-ai-backend[console]

# Docker sandbox support
pip install pydantic-ai-backend[docker]

# Everything
pip install pydantic-ai-backend[console,docker]

Quick Start — ConsoleCapability (Recommended)

The simplest way to give your agent filesystem tools:

from pydantic_ai import Agent
from pydantic_ai_backends import ConsoleCapability

agent = Agent("openai:gpt-4.1", capabilities=[ConsoleCapability()])

With Permissions

from pydantic_ai_backends import ConsoleCapability
from pydantic_ai_backends.permissions import READONLY_RULESET

# Read-only agent — write/edit/execute tools are hidden from the model
agent = Agent("openai:gpt-4.1", capabilities=[ConsoleCapability(permissions=READONLY_RULESET)])

Alternative: Toolset API

from dataclasses import dataclass
from pydantic_ai import Agent
from pydantic_ai_backends import LocalBackend, create_console_toolset

@dataclass
class Deps:
    backend: LocalBackend

agent = Agent(
    "openai:gpt-4.1",
    deps_type=Deps,
    toolsets=[create_console_toolset()],
)

backend = LocalBackend(root_dir="./workspace")
result = agent.run_sync(
    "Create a Python script that calculates fibonacci and run it",
    deps=Deps(backend=backend),
)
print(result.output)

That's it. Your agent can now:

• List files and directories (ls)
• Read and write files (read_file, write_file)
• Edit files with string replacement (edit_file)
• Search with glob patterns and regex (glob, grep)
• Execute shell commands (execute)

Available Backends

Backend	Storage	Execution	Use Case
StateBackend	In-memory	No	Testing, ephemeral sessions
LocalBackend	Filesystem	Yes	Local development, CLI tools
DockerSandbox	Container	Yes	Multi-user, untrusted code
CompositeBackend	Routed	Varies	Complex multi-source setups

In-Memory (StateBackend)

from pydantic_ai_backends import StateBackend

backend = StateBackend()
# Files stored in memory, perfect for tests

Local Filesystem (LocalBackend)

from pydantic_ai_backends import LocalBackend

backend = LocalBackend(
    root_dir="/workspace",
    allowed_directories=["/workspace", "/shared"],
    enable_execute=True,
)

Docker Sandbox (DockerSandbox)

from pydantic_ai_backends import DockerSandbox

sandbox = DockerSandbox(runtime="python-datascience")
sandbox.start()
# Fully isolated container environment
sandbox.stop()

Reusable Named Container

from pydantic_ai_backends import DockerSandbox

# Named container persists between sessions (packages survive restarts)
sandbox = DockerSandbox(
    image="python:3.12-slim",
    container_name="my-dev-env",  # implies auto_remove=False
    volumes={"/my/project": "/workspace"},
)
# Next time: finds existing container and reattaches

Console Toolset

Ready-to-use tools for pydantic-ai agents:

from pydantic_ai_backends import create_console_toolset

# All tools enabled
toolset = create_console_toolset()

# Without shell execution
toolset = create_console_toolset(include_execute=False)

# With approval requirements
toolset = create_console_toolset(
    require_write_approval=True,
    require_execute_approval=True,
)

# With custom tool descriptions
toolset = create_console_toolset(
    descriptions={
        "execute": "Run shell commands in the workspace",
        "read_file": "Read file contents from the workspace",
    }
)

Available tools: ls, read_file, write_file, edit_file, glob, grep, execute

Image Support

For multimodal models, enable image file handling:

toolset = create_console_toolset(image_support=True)

# Now read_file on .png/.jpg/.gif/.webp returns BinaryContent
# that multimodal models (GPT-4o, Claude, etc.) can see directly

Permission System

Fine-grained access control:

from pydantic_ai_backends import LocalBackend
from pydantic_ai_backends.permissions import DEFAULT_RULESET, READONLY_RULESET

# Safe defaults (allow reads, ask for writes)
backend = LocalBackend(root_dir="/workspace", permissions=DEFAULT_RULESET)

# Read-only mode
backend = LocalBackend(root_dir="/workspace", permissions=READONLY_RULESET)

Preset	Description
DEFAULT_RULESET	Allow reads (except secrets), ask for writes/executes
PERMISSIVE_RULESET	Allow most operations, deny dangerous commands
READONLY_RULESET	Allow reads only, deny all writes and executes
STRICT_RULESET	Everything requires approval

Docker Runtimes

Pre-configured environments:

Runtime	Base Image	Packages
python-minimal	python:3.12-slim	(none)
python-datascience	python:3.12-slim	pandas, numpy, matplotlib, scikit-learn
python-web	python:3.12-slim	fastapi, uvicorn, sqlalchemy, httpx
node-minimal	node:20-slim	(none)
node-react	node:20-slim	typescript, vite, react

Custom runtime:

from pydantic_ai_backends import DockerSandbox, RuntimeConfig

runtime = RuntimeConfig(
    name="ml-env",
    base_image="python:3.12-slim",
    packages=["torch", "transformers"],
)
sandbox = DockerSandbox(runtime=runtime)

Session Manager

Multi-user web applications:

from pydantic_ai_backends import SessionManager

# Docker (default)
manager = SessionManager(
    default_runtime="python-datascience",
    workspace_root="/app/workspaces",
)

# Each user gets isolated sandbox
sandbox = await manager.get_or_create("user-123")

Custom Sandbox Factory

Use any sandbox backend (Daytona, custom, etc.):

from pydantic_ai_backends import SessionManager, DaytonaSandbox

def daytona_factory(session_id: str) -> DaytonaSandbox:
    return DaytonaSandbox(sandbox_id=session_id)

manager = SessionManager(sandbox_factory=daytona_factory)
sandbox = await manager.get_or_create("user-123")

Why Choose This Library?

Feature	Description
Multiple Backends	In-memory, filesystem, Docker — same interface
Console Toolset	Ready-to-use tools for pydantic-ai agents
Permission System	Pattern-based access control with presets
Docker Isolation	Safe execution of untrusted code
Session Management	Multi-user support with workspace persistence
Image Support	Multimodal models can see images via BinaryContent
Pre-built Runtimes	Python and Node.js environments ready to go

Related Projects

Package	Description
Pydantic Deep Agents	Full agent framework (uses this library)
pydantic-ai-todo	Task planning toolset
subagents-pydantic-ai	Multi-agent orchestration
summarization-pydantic-ai	Context management
pydantic-ai	The foundation — agent framework by Pydantic

Contributing

git clone https://github.com/vstorm-co/pydantic-ai-backend.git
cd pydantic-ai-backend
make install
make test  # 100% coverage required

License

MIT — see LICENSE

---

Need help implementing this in your company?

We're Vstorm — an Applied Agentic AI Engineering Consultancy
with 30+ production AI agent implementations.

Made with ❤️ by Vstorm