inis-pydantic-ai

Pydantic AI toolset and capability for inis.run sandboxes.

Beta. Supported Pydantic AI range: >=1.71,<3 (1.71 is the first release with the capabilities API).

Install

pip install inis-pydantic-ai

Quickstart: inis_toolset()

Your app creates (or attaches to) the session and owns INIS_API_KEY; inis_toolset() just hands the model a bounded execution/file surface for it.

import asyncio

from inis import AsyncClient
from inis_pydantic_ai import inis_toolset
from pydantic_ai import Agent

async def main() -> None:
    client = AsyncClient()
    session = await client.sessions.create(egress_default="deny")
    toolset = inis_toolset(session_id=session.session_id)

    agent = Agent(
        "openai:gpt-4o",
        toolsets=[toolset],
        system_prompt="Run Python in the sandbox instead of guessing.",
    )

    try:
        result = await agent.run("What is 19 * 23? Use Python.")
        print(result.output)
    finally:
        await toolset.aclose()  # releases the toolset's local HTTP transport
        await session.destroy()  # this app owns the session; clean it up when done

asyncio.run(main())

session_id is required and inis_toolset() never creates, destroys, or discovers a session on its own — one agent thread maps to one session. To attach to a session another part of your app already owns instead of creating one, pass its ID the same way; this call never destroys it.

What it provides

Opt-in: session administration and one-shot sandboxes

Import from inis_pydantic_ai.lifecycle (not exported at the top level) only if you want the model itself to control these:

from inis_pydantic_ai.lifecycle import inis_ephemeral_toolset, inis_lifecycle_toolset

agent = Agent(
    "openai:gpt-4o",
    toolsets=[
        inis_toolset(session_id=session.session_id),
        inis_lifecycle_toolset(session_id=session.session_id),  # pause/resume/checkpoint/restore/expose/capture
    ],
)

InisSandbox capability

A capability (Agent(capabilities=[...])) following the pydantic-ai-harness sandbox contract — run_command, read_file, write_file, list_directory. Binds to one session for the life of the Agent instance rather than creating a fresh sandbox per run.

from inis import AsyncClient
from inis_pydantic_ai.sandbox import InisSandbox
from pydantic_ai import Agent

client = AsyncClient()
session = await client.sessions.create(egress_default="deny")

sandbox = InisSandbox(session=session)
agent = Agent("anthropic:claude-sonnet-5", capabilities=[sandbox])

try:
    result = await agent.run("Write a script that prints the first 10 primes and run it.")
    print(result.output)
finally:
    await sandbox.aclose()  # no-op here — `session` is host-owned
    await session.destroy()

Configuration

inis_toolset(
    session_id="ses_...",         # required
    api_key="inis_...",           # or $INIS_API_KEY
    base_url="https://api.inis.run",
)

Links

• Docs: docs.inis.run · Pydantic AI integration guide
• Source: github.com/inis-run/sdk/tree/main/pydantic-ai
• Homepage: inis.run

Built and run in the EU. Your code and data never leave Europe.

License

MIT — see LICENSE.