zygo — Python client

Run code you did not write — a customer's script, a plugin, something an LLM just generated — without letting it touch your host. Zygo keeps a warm sandbox per runtime, forks a fresh process for each request inside namespaces, cgroups and seccomp, and answers in about a millisecond. This package is the Python side of that: register a script once, then call it with JSON in and JSON out.

pip install zygo-sdk      # then `import zygo_sdk`

import zygo_sdk as zygo

client = zygo.connect()                 # `zygo api`, on loopback or a unix socket
resize = client.fn("resize")            # a function declared in sandbox.toml

out = resize({"url": "https://example.com/a.png"})
print(out.result, out.metrics.wall_ms)

Asynchronous, which is what an agent framework needs — the same methods, every one a coroutine:

import asyncio
import zygo_sdk as zygo

async def main(events):
    async with zygo.aio.connect() as client:
        return await asyncio.gather(*(client.call("resize", e) for e in events))

asyncio.run(main([{"url": "https://example.com/a.png"}]))

A refused request can be retried for you. Busy (the pool was full) and Unavailable (the host is still building or warming something) both mean the request never ran, so sending it again is safe; nothing else is retried:

client = zygo.connect(retries=3)        # waits the server's Retry-After, then again

A one-shot sandbox, needing nothing declared in advance:

r = client.run("python:3.12-slim", ["python3", "-c", "print(6 * 7)"], mem="128M")
print(r.stdout)

No dependencies. The standard library has an HTTP client, and a unix socket is thirty lines on top of it.

Full documentation: docs/book/17-api-sdk-mcp.md. The project: zygo.

Tests

python -m unittest discover -s tests

They run against a stand-in API and need no Linux, no kernel and no sandbox: what is under test is the client. A test that needs a real sandbox belongs in the Rust suites, against a real kernel.