ai2in 0.3.0

Python SDK for AI2IN — secure AI code sandboxes, hosted in India.

ai2in — Python SDK

The developer client for AI2IN.dev — secure AI code sandboxes, hosted in India. Spin up an isolated Linux sandbox, run untrusted / AI‑generated Python in a stateful kernel, and get back rich, typed results (text, HTML tables, charts, JSON) — the E2B code‑interpreter model, in‑region.

Stdlib‑only, zero dependencies.

Install

pip install ai2in            # once published
# or, from this repo:
pip install -e sdk/python

Quickstart

from ai2in import Sandbox

with Sandbox(api_key="ai2in_live_…", base_url="https://api.ai2in.dev") as sbx:
    execution = sbx.run_code(
        "import pandas as pd; pd.DataFrame({'gst_lakh_cr': [1.87, 1.73, 1.95]})"
    )

    print(execution.text)            # the main result's plain text
    print(execution.logs.stdout)     # ['…']

    for r in execution.results:      # rich, typed outputs
        if r.html:                   # e.g. a DataFrame → HTML table
            save(r.html)
        if r.png:                    # e.g. a matplotlib figure → base64 PNG
            save_image(r.png)

    if execution.error:              # structured error for self-correction
        print(execution.error.name, execution.error.value)

State persists across run_code calls in the same sandbox (like a notebook):

sbx.run_code("x = 41")
sbx.run_code("x + 1").text   # "42"

Streaming

Pass callbacks to stream output live as the kernel produces it (long loops, training logs, incremental prints). The events are still accumulated into the returned Execution, so the return value is identical whether or not you stream:

execution = sbx.run_code(
    "for i in range(5):\n    print('step', i)\n    import time; time.sleep(1)",
    on_stdout=lambda t: print(t, end="", flush=True),  # arrives chunk-by-chunk
    on_result=lambda r: render(r),                     # display() calls, figures, the value
    on_error=lambda e: print(e.value),
)
print(execution.execution_ms)                          # still get the full Execution

Result model

run_code() returns an Execution:

Field	Type	Notes
results	list[Result]	rich outputs; the cell's value has is_main_result=True
logs	Logs	.stdout / .stderr (lists of str)
error	ExecutionError | None	.name, .value, .traceback
text	str | None	convenience: the main result's text

Each Result may carry any of: text, html, markdown, svg, png, jpeg, pdf, latex, json, javascript. Call r.formats() for what's present.

Local dev

Point at a local engine (default http://localhost:4000):

with Sandbox() as sbx:            # talks to localhost:4000
    print(sbx.run_code("print(1 + 1)").logs.stdout)

---

© 2026 AI2IN.dev — see the repository LICENSE.