Isle Python SDK

Application environments for AI agents.

Install

Requires Python 3.9 or later.

python -m pip install isle-sdk

The package is named isle-sdk on PyPI and imported as isle. Get an API key from the Isle dashboard.

Version 0.2.0 adds persistent execution, capability discovery, and structured observations. Use a newly created environment and check its capabilities first; existing environments retain their original image.

The SDK automatically generates an idempotency key and reuses it for one transport retry. Pass idempotency_key= when the same create operation may be retried by a different process.

Quickstart

from isle import Client, IsleAPIError

isle = Client(api_key="isle_...")

# Create a KiCad sandbox
sandbox = isle.sandboxes.create(
    "kicad",
    name="Power supply board",
    idempotency_key="power-supply-board-v1",
)
sandbox.wait_until_ready()

# Upload a file
sandbox.files.upload("schematic.kicad_sch")

# Take a screenshot
image = sandbox.screen.screenshot()

# Control the mouse and keyboard
sandbox.mouse.click(500, 300)
sandbox.keyboard.type("100nF")
sandbox.keyboard.keypress(["CTRL", "S"])

# Download results
sandbox.files.download("/home/user/work/board.kicad_pcb")

# Stop (archives the sandbox, can be resumed later)
sandbox.stop()
sandbox.wait_until_stopped()

# Resume a stopped sandbox
sandbox.resume()

# Lifecycle failures are returned by the API and retained on the environment
sandbox.refresh()
if sandbox.last_error:
    print(sandbox.last_error, sandbox.last_error_at)

# Retrieve timestamped error history (newest first)
for error in sandbox.errors():
    print(error["source"], error["message"], error["created_at"])

# Recording metadata distinguishes a complete recording from the bounded
# 512 MiB / six-hour / low-disk partial recording policy.
recording = sandbox.recording_info()
print(recording["url"], recording["truncated"], recording["truncation_reason"])

# Permanently delete a stopped sandbox and its retained data
sandbox.stop()
sandbox.wait_until_stopped()
sandbox.destroy(timeout=120)

API failures raise IsleAPIError. Its status_code, error, and metadata attributes can be used to handle failures without parsing an exception string. For example, a provisioning failure includes its recoverable metadata["environment_id"] when the API created a record before the remote environment failed to start.

Environments

• kicad - EDA environment with KiCad
• freecad - CAD environment with FreeCAD
• blender - 3D environment with live bpy execution, native .blend files and asset-preserving checkpoints

See the Blender guide and example for inspection, live edits, saving and recovery.

Persistent execution

Use Isle SDK 0.2.0 or later with a newly created KiCad or FreeCAD environment. Existing environments keep their original image. The API is model-neutral and works with any agent framework or HTTP client.

python -m pip install --upgrade 'isle-sdk>=0.2.0'

Connect to an existing environment running the updated image:

import uuid
from isle import Client, ExecutionSubmissionUnknown

client = Client()  # ISLE_API_KEY stays in the client
sandbox = client.sandboxes.get("YOUR_SANDBOX_ID")
capabilities = sandbox.capabilities()
assert capabilities["execution"]["available"]
generation = capabilities["generation"]
execution_id = str(uuid.uuid4())  # preserve this and generation before POST

try:
    result = sandbox.execution.submit(
        "counter = 41\nlog(counter)\ndisplay(pyautogui.screenshot())",
        execution_id=execution_id,
        expected_generation=generation,
        timeout_seconds=30,
    )
except ExecutionSubmissionUnknown as error:
    result = sandbox.execution.get(error.execution_id, generation=error.generation)

if result["status"] == "running":
    result = sandbox.execution.wait(execution_id, generation=generation)
print(result["status"], result["stdout"], result["error"])
observation = sandbox.screen.observe()

execution.run() combines submit and wait. execution.cancel(execution_id, expected_generation=generation) cancels a running cell. ExecutionWaitTimeout preserves the ID and generation and does not cancel remote work. After any unknown outcome, query the original execution; do not replay under a new ID. A missing or expired record requires inspecting the desktop and saved files.

The default namespace includes pyautogui, time, log(), and display(). Combine Python imports, short GUI actions, installed CLI tools, and project files beneath /home/user/work; keep PyAutoGUI's fail-safe enabled. Variables persist between cells, while Python descendants pause between them. One execution or other desktop operation runs at a time; an explicit desktop_busy rejection can be retried with bounded backoff and the same request identity when the API confirms it was not started or was rejected.

Cancellation and server timeouts verify worker and held-key/button cleanup, then reset the namespace. Completed edits remain. Results retain their original generation and expose namespace_reset; refresh capabilities before new work. A Python exception also returns a result, so inspect status and error rather than treating HTTP success as successful code execution.

Current cells allow 64 KiB of UTF-8 source and 1–120 seconds. stdout and stderr are bounded to 64 KiB each. At most eight emitted outputs share a 1 MiB budget for PNG bytes and text; execution JSON is limited to 2 MiB. Read capabilities["limits"] and output_truncated. PNGs from display() are base64 image outputs in bounded runtime records, without durable image URLs. Download them while available or save them to the workspace. Observations carry a PNG, timestamp, generation, dimensions, scale factor, and active_document=None; retained observations use screenshot storage.

See the execution reference for the HTTP contract. Keep model-provider credentials in your client process.

Shell bridge for an existing agent session

From an authorized repository checkout, Codex or any agent with shell access can use examples/codex/desktop.py with Isle SDK 0.2.0 or later and an existing execution-enabled environment:

python examples/codex/desktop.py --sandbox-id SANDBOX_ID capabilities
python examples/codex/desktop.py --sandbox-id SANDBOX_ID observe
python examples/codex/desktop.py --sandbox-id SANDBOX_ID run action.py --generation GENERATION

Set ISLE_API_KEY in the host environment. The bridge makes no model API calls and requires no model-provider SDK or separate model API key. It records request IDs before writes and returns local PNG, text, and result paths for the agent to inspect. See examples/codex/README.md in that checkout for polling, cancellation, downloads, and independent application verification. The bridge is a repository example, not an installed isle-sdk command.

Bounded KiCad and FreeCAD mechanical tasks have been validated through Codex using the deployed API, independent saved-file checks, and visual review.

Astra example

After installing the SDK above, add the optional client dependencies:

python -m pip install -r examples/astra/requirements.txt
python examples/astra/run.py YOUR_SANDBOX_ID

examples/astra/run.py connects the OpenAI Responses API from your client process. Keep OPENAI_API_KEY there, outside desktop scripts and files, and pass emitted PNGs to Astra as image inputs. examples/astra/validate.py checks saved KiCad copies with CLI ERC, DRC, and schematic parity. New KiCad IPC images expose the official kipy client through execution cells; check for kicad_ipc in capabilities and see examples/kicad/README.md in the source checkout. An Isle MCP connector is not included. This optional Responses API adapter has offline coverage; its own live API run remains pending. The Codex-driven application checks above do not establish that adapter's live behavior. See the integration guide and execution reference.

Development

From an authorized Isle repository checkout:

python -m pip install -e ./sdk/python
python -m unittest discover -s sdk/python/tests