Pynox

A Python with the sandbox built in. Made for the code your agents write.

Start here

Pynox is a Python. One binary carries a complete CPython, so pynox app.py runs the same file python app.py runs, with the same output and exit code. Around the interpreter, Pynox adds three things, in the order you will meet them:

1. A sandbox on every run. The script reads its own directory and the standard library. Everything else needs a grant: a flag for one run, a pynox.toml for a project. When the script oversteps, you get Python's traceback plus one line that says which grant would fix it.
2. A pool for agents. Code written by a model runs in a fresh process forked from a warm template, under the permissions you chose, in about 5 ms, and comes back as a result the model can act on.
3. Optimizations that need no flag. Pynox speeds up specific operations and publishes the measurements. Today: json.dump, 1.9× on a 3.9 MiB file.

--no-sandbox is the plain interpreter whenever you need it.

What you can build

Each one is a runnable example tested on every commit. Real agents shows the same patterns with Pydantic AI, Strands Agents and Claude, copied from real runs: five Python runs for a full analysis cost 7 ms inside Pynox.

Three ways to run

pynox report.py                 # your script, sandboxed, output and exit code are Python's
pynox --json report.py          # the same run as one JSON result, for automation
pynox --no-sandbox -O report.py # CPython's own command line, unrestricted

For a model's code, pynox.tool() from the pynox package returns a function any agent framework registers as a tool; pynox.run(code) does the same from plain Python. The agents guide shows both.

Why it exists

An agent answers a question by writing Python and running it. Until now that code ran with your privileges, where one ambiguous instruction becomes rmtree on the wrong directory, or in a remote container that costs a second and a bill per call. Pynox puts kernel isolation between the model and your machine and keeps the whole Python ecosystem on the model's side of it.

Guides

Agents

Runtime

Project