Onyx Python Sandbox
https://github.com/onyx-dot-app/python-sandbox- Category
- AI Tools
- Rank
- No. 1492Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- onyx-dot-app
- GitHub
- 28 stars
- Latest release
- code-interpreter-0.4.6
- Date
About
Secure and lightweight Python code execution environment for LLMs.
What it does
Onyx runs submitted Python inside disposable containers, then returns output, errors, status, timing, and created files through a REST service. It supports uploaded inputs, streamed results, notebook-like display of a final expression, and longer-lived sessions for related commands.
Why it's ranked here
This is a strong, practical interpreter service because isolation is visible in the implementation, not merely promised. Container processes lose most capabilities, run without privilege escalation, receive memory and process limits, and default to no network access. Docker and Kubernetes backends broaden its operational fit, while end-to-end tests cover execution and file artifacts.
What's good
Resource controls cover wall-clock timeout, CPU time, memory, process count, and output size. Relative-path validation rejects traversal, workspaces use temporary storage, and generated files return through managed identifiers. The API offers synchronous and streamed execution, health reporting, file exchange, and expiring sessions. Network access is disabled by default.
Tradeoffs
The recommended Docker setup mounts the host daemon socket and runs the service as root, so it belongs only in a trusted environment. The alternative requires a privileged container. Kubernetes network lockdown uses a privileged initialization step with network administration capability unless operators disable it. File storage is local, UUID-based disk storage rather than an external durable store.
How to use it well
Use it when an agent needs local Python analysis, plots, file transformations, or several commands within an expiring workspace. Choose Docker for simpler trusted deployments and Kubernetes for pod-based operation. Set conservative resource limits and preserve the default network block. It does not provide hosted execution, so you must operate the service and container infrastructure.
Technical notes+
code-interpreter/app/main.py builds the FastAPI service, exposes health and version metadata, pulls the Docker executor image during startup, and runs a 30-second expired-session reaper. code-interpreter/app/api/routes.py implements synchronous execution, Server-Sent Events streaming, file operations, sessions, and session Bash execution. code-interpreter/app/services/executor_factory.py selects Docker or Kubernetes. code-interpreter/app/services/executor_docker.py applies a 64-process limit, no-new-privileges, dropped capabilities with CHOWN restored, tmpfs workspaces, optional CPU and memory limits, and a default network selected in code-interpreter/app/app_configs.py as none. code-interpreter/app/services/executor_kubernetes.py runs as UID 65532, drops capabilities, uses bounded emptyDir volumes, and defaults to an init container that blocks IPv4 and IPv6 egress with NET_ADMIN. Its exit-code parser uses eval on an exec error-channel string, which deserves security scrutiny. code-interpreter/app/services/file_storage.py stores file bytes and JSON metadata on local disk. code-interpreter/tests/e2e/test_basic_flow.py exercises health, basic execution, edited inputs, generated plots, and multiple output files.
Observed
- License
- MIT License
- Primary language
- Python
- Interface
- FastAPI REST API with OpenAPI documentation and Server-Sent Events streaming
- Packaging and install surface
- Published Docker image with Docker-out-of-Docker and Docker-in-Docker launch options
- Platform support
- Docker and Kubernetes executor backends; Helm and Kubernetes deployment instructions are referenced
- Execution model
- Disposable isolated containers plus expiring long-lived sessions
- Tests
- Repository includes end-to-end tests for execution, uploaded files, generated PNG output, and multiple output files
Read from README.md, code-interpreter/app/main.py, code-interpreter/app/image_ref.py, code-interpreter/app/app_configs.py, code-interpreter/app/logging_config.py, code-interpreter/app/api/routes.py, code-interpreter/app/models/schemas.py, code-interpreter/app/services/file_storage.py, code-interpreter/app/services/executor_base.py, code-interpreter/app/services/executor_docker.py, code-interpreter/app/services/executor_factory.py, code-interpreter/app/services/executor_kubernetes.py, code-interpreter/tests/e2e/test_basic_flow.py, code-interpreter/tests/integration_tests/conftest.py.
What it can do
Execute Python code safely
Python code or script → Execution results and output
Validate Python syntax
Python code string → Syntax validation results
Isolate code execution
Potentially unsafe Python code → Sandboxed execution environment
Process LLM-generated Python code
Python code from language models → Safe execution results
Run Python scripts with resource limits
Python script → Controlled execution with safety constraints
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.