API Reference

Sandbox Class

The main class for creating and managing sandboxes.

Class Methods

Sandbox.create()

Creates a new sandbox instance.

@classmethod
def create(
    variant: str = "datalayer",
    timeout: float | None = None,
    environment: str | None = None,
    gpu: str | None = None,
    cpu: float | None = None,
    memory: int | None = None,
    env: dict[str, str] | None = None,
    tags: dict[str, str] | None = None,
    name: str | None = None,
    network_policy: str | None = None,
    allowed_hosts: list[str] | None = None,
    examples: list[tuple[str, str]] | None = None,
    config: SandboxConfig | None = None,
    **kwargs,
) -> Sandbox

Parameters:

Parameter	Type	Description
variant	str	Sandbox type: "cloudflare", "coreweave", "datalayer", "daytona", "docker", "e2b", "eval", "google-colab", "jupyter-server", "kaggle", "modal", or "monty". Defaults to "datalayer".
timeout	float	Execution timeout in seconds
environment	str	Runtime environment name
gpu	str	GPU type (e.g., "T4", "A100", "H100")
cpu	float	Number of CPU cores
memory	int	Memory in MB
env	dict	Environment variables
tags	dict	Metadata tags
name	str	Sandbox name
network_policy	str	Network access policy: "inherit", "none", "allowlist", "all"
allowed_hosts	list[str]	Hosts reachable when the policy is "allowlist"
examples	list[tuple[str, str]]	(title, code) pairs the REPL lists under :examples
config	SandboxConfig	Full configuration object
**kwargs	dict	Variant-specific arguments handed to the sandbox constructor — template= (e2b), container_image= (coreweave), api_url= (cloudflare), snapshot_name= (datalayer), and the rest documented on each variant's page

Sandbox.from_id()

Reconnects to an existing sandbox.

@classmethod
def from_id(sandbox_id: str, **kwargs) -> Sandbox

Datalayer only — it reconnects to a Datalayer runtime whatever else is running. **kwargs are that variant's: token=, run_url=.

Sandbox.list()

Lists running sandboxes, as sandbox objects.

@classmethod
def list(tags: dict[str, str] | None = None, **kwargs) -> Iterator[Sandbox]

Datalayer only. For any other variant, ask its manager: get_manager(variant).list(), which answers with list[SandboxInfo] — see the CLI page.

Sandbox.list_environments()

Lists available environments for a sandbox variant.

@classmethod
def list_environments(
    variant: str = "datalayer",
    **kwargs,
) -> list[SandboxEnvironment]

Parameters:

Parameter	Type	Description
variant	str	Sandbox type: "cloudflare", "coreweave", "datalayer", "daytona", "docker", "e2b", "eval", "google-colab", "jupyter-server", "kaggle", "modal", or "monty"
**kwargs	dict	Variant-specific arguments (e.g., credentials, run URL)

Legacy local-eval, local-docker, and local-jupyter variant names are not supported.

Instance Methods

run_code()

Executes Python code in the sandbox.

def run_code(
    code: str,
    language: str = "python",
    context: Context | None = None,
    on_stdout: Callable[[OutputMessage], None] | None = None,
    on_stderr: Callable[[OutputMessage], None] | None = None,
    on_result: Callable[[Result], None] | None = None,
    on_error: Callable[[CodeError], None] | None = None,
    envs: dict[str, str] | None = None,
    timeout: float | None = None,
) -> ExecutionResult

run_code_streaming()

Executes Python code and yields output/result/error events as they arrive.

def run_code_streaming(
    code: str,
    language: str = "python",
    context: Context | None = None,
    envs: dict[str, str] | None = None,
    timeout: float | None = None,
) -> Iterator[OutputMessage | Result | CodeError]

run_code_streaming_async()

Async version of run_code_streaming().

async def run_code_streaming_async(
    code: str,
    language: str = "python",
    context: Context | None = None,
    envs: dict[str, str] | None = None,
    timeout: float | None = None,
) -> AsyncIterator[OutputMessage | Result | CodeError]

start()

Starts the sandbox.

def start() -> None

stop()

Stops the sandbox gracefully.

def stop() -> None

terminate() / kill()

Datalayer only; both are aliases for stop(). Every other variant is stopped with stop() — or by leaving its with block.

def terminate() -> None
def kill() -> None

set_timeout()

Extends the life of a running sandbox. Only two variants offer it, and the argument is not spelled the same in both.

def set_timeout(timeout_seconds: float) -> None  # datalayer
def set_timeout(seconds: float) -> None          # e2b

Properties

Property	Type	Description
sandbox_id	str | None	Unique sandbox identifier — None until start() has run
files	SandboxFilesystem	Filesystem interface
commands	SandboxCommands	Command execution interface
tags	dict[str, str]	Metadata tags

---

CodeSandboxClient

CodeSandboxClient is a high-level, variant-agnostic facade over a sandbox. It normalizes one-shot outcomes and streaming events across all variants.

Methods

execute_code()

def execute_code(
    code: str,
    language: str = "python",
    timeout: float | None = None,
    envs: dict[str, str] | None = None,
) -> CodeExecutionOutcome

execute_code_async()

async def execute_code_async(
    code: str,
    language: str = "python",
    timeout: float | None = None,
    envs: dict[str, str] | None = None,
) -> CodeExecutionOutcome

execute_code_streaming()

def execute_code_streaming(
    code: str,
    language: str = "python",
    timeout: float | None = None,
    envs: dict[str, str] | None = None,
) -> Iterator[OutputMessage | Result | CodeError]

execute_code_streaming_async()

async def execute_code_streaming_async(
    code: str,
    language: str = "python",
    timeout: float | None = None,
    envs: dict[str, str] | None = None,
) -> AsyncIterator[OutputMessage | Result | CodeError]

Compatibility-free execution facade

Consumers that need a Jupyter-shaped reply can stay on the sandbox client without accessing a variant's underlying kernel implementation:

reply = client.execute(code, timeout=60)
reply = client.execute_interactive(code, output_hook=handle_output)

The client also exposes variant-neutral lifecycle and variable operations:

client.start()
client.interrupt()
client.restart()
client.set_variable("name", value)
client.set_variables({"one": 1, "two": 2})
value = client.get_variable("name")
client.stop()

Properties

Property	Description
id	Active backend identifier, including a remote kernel ID when applicable
variant	Configured sandbox variant
config	Variant-neutral sandbox configuration
info	Runtime sandbox information after startup
kernel_info	Language metadata without exposing the underlying kernel client
is_started	Whether the sandbox has been started

Outcome model

CodeExecutionOutcome includes normalized fields such as:

• success
• execution_ok
• stdout
• stderr
• results
• execution_error
• code_error
• exit_code
• interrupted

---

Kaggle and Colab Clients

Code Sandboxes owns the provider-specific clients used by its managed notebook variants. They are exported from code_sandboxes:

from code_sandboxes import (
    GoogleColabKernelClient,
    KaggleExecutionResult,
    KaggleKernelClient,
    KaggleKernelExecutor,
    parse_google_colab_channels_url,
    parse_kaggle_channels_url,
)

• KaggleKernelClient connects to an interactive Kaggle notebook kernel.
• KaggleKernelExecutor submits and monitors Kaggle batch notebook jobs.
• KaggleExecutionResult normalizes batch output into a Jupyter-like reply.
• GoogleColabKernelClient connects to an already-running Google Colab kernel.
• The parser helpers extract connection details from browser channels URLs.

See Kaggle and Google Colab for authentication and complete examples.

---

SandboxFilesystem

File operations interface.

Methods

read()

def read(path: str) -> str

Reads file contents as a string.

read_bytes()

def read_bytes(path: str) -> bytes

Reads file contents as bytes.

write()

def write(path: str, content: str, make_dirs: bool = True) -> None

Writes text to a file, creating parent directories unless told not to.

write_bytes()

def write_bytes(path: str, content: bytes, make_dirs: bool = True) -> None

Writes bytes to a file. write() handles text only.

list()

def list(path: str = "/") -> list[FileInfo]

Lists directory contents.

mkdir()

def mkdir(path: str, parents: bool = True) -> None

Creates a directory.

rm()

def rm(path: str, recursive: bool = False) -> None

Removes a file or directory.

exists() / is_file() / is_dir()

def exists(path: str) -> bool
def is_file(path: str) -> bool
def is_dir(path: str) -> bool

Asks about a path without reading it.

copy() / move()

def copy(src: str, dst: str) -> None
def move(src: str, dst: str) -> None

get_info()

def get_info(path: str) -> FileInfo

One entry's metadata, without listing its directory.

upload()

def upload(local_path: str, remote_path: str) -> None

Uploads a local file to the sandbox.

download()

def download(remote_path: str, local_path: str) -> None

Downloads a file from the sandbox.

FileInfo

What list() and get_info() answer with.

@dataclass
class FileInfo:
    name: str
    path: str
    type: FileType = FileType.FILE   # FILE, DIRECTORY, SYMLINK
    size: int = 0                    # bytes, for files
    modified: float = 0.0            # Unix timestamp
    permissions: str = ""

    # Convenience properties
    @property
    def is_file(self) -> bool: ...
    @property
    def is_directory(self) -> bool: ...

---

SandboxCommands

Command execution interface.

Methods

run()

def run(
    command: str,
    cwd: str | None = None,
    env: dict[str, str] | None = None,
    timeout: float | None = None,
    shell: bool = True,
) -> CommandResult

Runs a command and waits for completion.

exec()

def exec(*args: str, **kwargs) -> ProcessHandle

Executes a command with streaming output.

spawn()

def spawn(command: str, **kwargs) -> ProcessHandle

Starts a background process.

CommandResult

@dataclass
class CommandResult:
    exit_code: int = 0
    stdout: str = ""
    stderr: str = ""
    duration: float = 0.0

    @property
    def success(self) -> bool: ...  # exit_code == 0

ProcessHandle

A process still running. stdout and stderr are iterators that yield as output arrives.

@dataclass
class ProcessHandle:
    command: str
    pid: int | None = None

    @property
    def stdout(self) -> Iterator[str]: ...
    @property
    def stderr(self) -> Iterator[str]: ...
    @property
    def returncode(self) -> int | None: ...

    def read_stdout(self) -> str: ...
    def read_stderr(self) -> str: ...
    def wait(self, timeout: float | None = None) -> int: ...
    def poll(self) -> int | None: ...
    def terminate(self) -> None: ...
    def kill(self) -> None: ...

---

Data Models (Pydantic)

ExecutionResult

Complete result of a code execution with detailed status information.

class ExecutionResult(BaseModel):
    # Results and logs
    results: list[Result]
    logs: Logs

    # Execution-level (infrastructure) status
    execution_ok: bool = True  # Did sandbox infrastructure work?
    execution_error: str | None = None  # Infrastructure failure details

    # Code-level (user code) status
    code_error: CodeError | None = None  # Python exception info

    # Process exit status (sys.exit)
    exit_code: int | None = None  # Exit code if code calls sys.exit()

    # Metadata
    execution_count: int = 0
    context_id: str | None = None
    started_at: float | None = None
    completed_at: float | None = None
    interrupted: bool = False

    # Properties
    @property
    def success(self) -> bool:
        """True if execution_ok and no code_error and not interrupted and exit_code is 0 or None"""

    @property
    def duration(self) -> float | None:
        """Execution duration in seconds"""

    @property
    def text(self) -> str | None:
        """Main text result"""

    @property
    def stdout(self) -> str:
        """All stdout as single string"""

    @property
    def stderr(self) -> str:
        """All stderr as single string"""

Usage:

result = sandbox.run_code("x = 1 + 1; print(x)")

# Check infrastructure success
if not result.execution_ok:
    print(f"Sandbox failed: {result.execution_error}")
# Check explicit process exit
elif result.exit_code not in (None, 0):
    print(f"Process exited with code: {result.exit_code}")

# Check code-level error
elif result.code_error:
    print(f"Python error: {result.code_error.name}: {result.code_error.value}")

# Success!
else:
    print(f"Result: {result.text}")
    print(f"Duration: {result.duration:.2f}s")

# Or use convenience property
if result.success:
    print("Everything worked!")

CodeError

Error information when Python code raises an exception.

class CodeError(BaseModel):
    name: str          # Error class name (e.g., "ValueError")
    value: str         # Error message
    traceback: str = ""  # Full traceback

SandboxEnvironment

class SandboxEnvironment(BaseModel):
    name: str
    title: str
    language: str = "python"
    owner: str = "local"
    visibility: str = "local"
    burning_rate: float = 0.0
    metadata: dict[str, Any] | None = None
    cpu: str | None = None
    memory: str | None = None
    gpu: str | None = None
    gpu_count: int | None = None
    gpu_memory: int | None = None

Result

A single execution result with multiple representations.

class Result(BaseModel):
    data: dict[str, Any]      # MIME type to content mapping
    is_main_result: bool = False
    extra: dict[str, Any]     # Additional metadata

    # Convenience properties
    @property
    def text(self) -> str | None: ...  # text/plain
    @property
    def html(self) -> str | None: ...  # text/html
    @property
    def json(self) -> Any: ...         # application/json
    @property
    def png(self) -> str | None: ...   # image/png (base64)
    @property
    def jpeg(self) -> str | None: ...  # image/jpeg (base64)
    @property
    def svg(self) -> str | None: ...   # image/svg+xml
    @property
    def markdown(self) -> str | None: ...  # text/markdown

OutputMessage

class OutputMessage(BaseModel):
    line: str
    timestamp: float = 0.0
    error: bool = False       # True for stderr

Logs

class Logs(BaseModel):
    stdout: list[OutputMessage]
    stderr: list[OutputMessage]

    @property
    def stdout_text(self) -> str: ...  # All stdout as string
    @property
    def stderr_text(self) -> str: ...  # All stderr as string

Context

class Context(BaseModel):
    id: str
    language: str = "python"
    cwd: str | None = None
    env: dict[str, str]

SandboxConfig

class SandboxConfig(BaseModel):
    timeout: float = 30.0
    memory_limit: int | None = None
    cpu_limit: float | None = None
    environment: str = "ai-agents-env"
    working_dir: str | None = None
    env_vars: dict[str, str]
    gpu: str | None = None
    name: str | None = None
    network_policy: str = "inherit"  # "inherit", "none", "allowlist", "all"
    allowed_hosts: list[str]
    idle_timeout: float | None = None
    max_lifetime: float = 86400.0  # 24 hours
    examples: list[tuple[str, str]]  # (title, code) pairs for the REPL

SandboxInfo

class SandboxInfo(BaseModel):
    id: str
    variant: str
    status: SandboxStatus = SandboxStatus.RUNNING
    created_at: float = 0.0
    end_at: float | None = None
    config: SandboxConfig | None = None
    name: str | None = None
    metadata: dict[str, Any]
    resources: ResourceConfig | None = None

    @property
    def remaining_time(self) -> float | None: ...  # Seconds until termination

SnapshotInfo

What create_snapshot() and list_snapshots() answer with — datalayer only.

class SnapshotInfo(BaseModel):
    id: str
    name: str
    sandbox_id: str
    created_at: float = 0.0
    size: int = 0
    description: str = ""

JupyterServerOptions

What prepare_jupyter_server() is asked for — see Jupyter over Provider Ingress.

class JupyterServerOptions(BaseModel):
    port: int = 8888
    token: str | None = None          # generated when omitted
    install_if_missing: bool = True
    install_timeout: float = 180.0

JupyterServerEndpoint

What it answers with. headers authenticate the provider ingress and query authenticates Jupyter itself; both are secrets, and both are kept out of the repr.

class JupyterServerEndpoint(BaseModel):
    port: int
    http_url: str
    websocket_url: str
    headers: dict[str, str]
    query: dict[str, str]

ResourceConfig

class ResourceConfig(BaseModel):
    cpu: float | None = None         # CPU cores
    memory: int | None = None        # Memory in MB
    gpu: str | None = None           # GPU type ("T4", "A100", etc.)
    gpu_count: int = 1               # Number of GPUs
    disk: int | None = None          # Disk in GB

---

Management API

The verbs the CLI runs on, from Python. A manager exists per variant, and states what it can do rather than failing late.

from code_sandboxes import get_manager, manageable_variants

manageable_variants()            # the variants a manager exists for
manager = get_manager("modal")
manager.capabilities             # frozenset of the verbs this variant answers
manager.list()                   # list[SandboxInfo]
manager.get(sandbox_id)          # SandboxInfo | None
manager.create(**kwargs)         # SandboxInfo
manager.update(sandbox_id, **changes)
manager.delete(sandbox_id)       # bool

A verb outside capabilities raises SandboxManagementError. See the CLI page for what each variant answers.

Provider Registry

What a variant needs before it can run anything — credentials, and the extra that installs it. It reports what is present; it starts nothing and reads no secret's value.

from code_sandboxes import PROVIDERS, available_providers, get_provider

PROVIDERS                        # every provider declared
available_providers()            # those whose credentials are on this machine
get_provider("daytona")          # one, or None

---

Exceptions

Exception	Description
SandboxError	Base exception for all sandbox errors
SandboxTimeoutError	Execution timed out
SandboxExecutionError	Code execution failed
SandboxNotStartedError	The sandbox was used before start()
SandboxConnectionError	Connection to sandbox failed
SandboxConfigurationError	The variant cannot be used as configured — a missing extra, or a setting it refuses
ContextNotFoundError	No such execution context
VariableNotFoundError	No such variable in the context
SandboxSnapshotError	Snapshot operation failed
SandboxResourceError	Resource limit exceeded
SandboxAuthenticationError	Authentication failed
SandboxQuotaExceededError	Usage quota exceeded