Independent project: Cageforge is not affiliated with, sponsored by, or endorsed by OpenAI.

Cageforge Python binding

cageforge provides a typed Python API for the Cageforge sandbox on Linux, macOS, and Windows. It loads the selected profile and uses the native backend provided by the installed wheel.

The package is built with PyO3 and maturin. The published package name is cageforge.

Use the published cageforge crate as the project-level reference for the sandbox API. See the shared configuration guide for the complete Linux/macOS/Windows profile shape, including the resources a first restricted launch must declare separately from local IPC.

Install

python -m pip install cageforge

Each wheel contains the Python extension and the native resources for one operating-system and architecture target. Linux wheels include the pinned Bubblewrap resource. Applications do not need to install a separate native helper or provide a host-specific resource path.

The release also includes free-threaded CPython 3.14 wheels for each supported operating-system and architecture target. Standard CPython 3.10 and newer builds use the stable abi3 wheel for their platform.

Quick start with the repository smoke profile

The repository already contains one runnable profile for each supported OS. This example selects the profile for the current host and uses Cageforge.from_toml_file to run it. Run it from the repository root:

import platform
from pathlib import Path

from cageforge import Cageforge, PermissionApprover, RuntimeContext

platform_name = platform.system().lower()
profile_name = {
    "linux": "linux",
    "darwin": "macos",
    "windows": "windows",
}[platform_name]
profile = (
    Path("crates")
    / "cageforge-config"
    / "examples"
    / "runnable"
    / profile_name
    / "smoke.toml"
).resolve()

context = RuntimeContext(profile.parent)
toml = profile.read_bytes().decode("utf-8")
Cageforge.check_toml(toml, context=context)
request = Cageforge.permission_request(toml, context=context)
grant = PermissionApprover().approve(request)
with Cageforge.from_toml_file(
    profile, context=context, grant=grant, request=request
) as runtime:
    with runtime.launch() as process:
        print(process.read_stdout(4096).decode().strip())
        assert process.wait().exit_code == 0

The three profiles are linux/smoke.toml, macos/smoke.toml, and windows/smoke.toml. They use the current Cageforge TOML schema, include minimal read access, declare a workspace root, allow writes to workspace-root, and disable the network. The Windows profile uses cmd.exe; the POSIX profiles use /bin/echo.

Profiles with approval.mode = "preflight" require a trusted host grant before launch. PermissionRequest is descriptive; only PermissionApprover can issue the opaque PermissionGrant. A grant never changes an already-running process. When permission_request is called with custom identity or digest arguments, pass that same PermissionRequest to from_toml or from_toml_file; the runtime then authorizes the grant against the exact identity that was approved.

PermissionRequest.local_ipc() returns frozen typed LocalIpcEndpoint values. Their kind is unix_socket or windows_named_pipe, and value is the validated native endpoint. The platform overlay selects which endpoint kind is present; a Windows named-pipe request remains fail-closed if the native backend cannot prove the required isolation.

Persistent grants and store paths

The permission store is host state, not TOML policy. Choose its absolute path explicitly and use a persistent grant when the approval should survive a new process:

from pathlib import Path

from cageforge import PermissionApprover, PermissionStore

store = PermissionStore.open(Path("/var/lib/my-tool/permissions.json"))
toml = profile.read_bytes().decode("utf-8")
request = Cageforge.permission_request(toml, context=context)
grant = PermissionApprover().approve(request, scope="persistent")
store.put(grant, request)

cached = store.get(request)
assert cached is not None
with Cageforge.from_toml_file(
    profile, context=context, grant=cached, request=request
) as runtime:
    ...

PermissionStore protects the file with owner-only permissions on Unix and an owner-only DACL on Windows. It uses a versioned JSON document, a kernel file lock, and atomic replacement. A missing store record is not an approval; preflight remains deny-by-default.

Persistent grants can be inspected and revoked by stable ID without exposing their approved capability payload:

page = store.list_page(page_size=50)
for summary in page.entries:
    print(summary.id, summary.tool_id)
if page.next_cursor is not None:
    page = store.list_page(page_size=50, cursor=page.next_cursor)

result = store.revoke(request.grant_id())
assert str(result) in {"revoked", "not-found"}

GrantPageCursor is opaque and becomes stale when the store changes. The binding exposes stable store exception subclasses for invalid IDs, invalid cursors, page size, stale cursors, lock, read, write, and format failures. All digest and store validation is performed by the shared Rust implementation.

When profile_name is omitted, Cageforge.from_toml and check_toml use the TOML document's default_profile. Cageforge.from_toml_file reads a file and uses its parent directory as the default current directory. RuntimeContext() uses the Python process directory and lets the native adapter provide the platform's default minimal paths. Passing a path in RuntimeContext does not grant access unless the selected profile contains the corresponding rule.

The minimal selector is symbolic. Linux adapters supply the executable and loader paths needed by the selected backend, macOS supplies its system runtime paths, and Windows supplies the system root and System32 paths. Use the separate files under cageforge-config/examples/runnable/ when the command or path syntax is OS-specific.

Processes, asyncio, and errors

SandboxProcess provides try_wait, wait, wait_for, kill, and close, as well as read_stdout, read_stderr, write_stdin, and close_stdin. The wait_for method is a compatibility alias for wait. Blocking native waits and stream operations release the GIL.

For asyncio applications, wait_for_async(process) waits in a worker thread so the event loop remains responsive. Cancelling the coroutine terminates the process boundary and then propagates asyncio.CancelledError.

Configuration, initialization, launch, process, stream, and Windows setup failures use typed exceptions under CageforgeError, including CageforgeConfigurationError, CageforgeLaunchError, and CageforgeProcessError:

from cageforge import Cageforge, CageforgeConfigurationError

try:
    Cageforge.check_toml("not valid = [")
except CageforgeConfigurationError as error:
    print(f"invalid Cageforge configuration: {error}")

Windows setup and native resources

Windows provisioning is explicit because installation can require UAC. Before launching on Windows, an application can reconcile and verify the owner-scoped setup:

from cageforge import WindowsSetup

if WindowsSetup.is_supported():
    if WindowsSetup.status() != "ready":
        WindowsSetup.install()
    WindowsSetup.verify()

install() is the only operation in this sequence that may request elevation. Creating a runtime does not silently install Windows components. Linux wheels prefer a compatible system Bubblewrap and fall back to the bundled Bubblewrap resource shipped in the wheel. macOS uses the packaged native helper.

Development

From this directory, install maturin and the Python development tools, then use maturin develop, pytest, mypy --strict, and ruff check. The committed _cageforge.pyi stub is generated by running cargo run --bin stub_gen from this crate directory. The Java and Python bindings share the same native policy contract; binding-specific contract checks live with their respective tests.