Pydantic AI

Pydantic AI is the recommended Python agent framework for new projects. The inis integration wraps the Python SDK as a ready-made toolset — persistent session, shell, files, and code execution.

Install

pip install inis-pydantic-ai

Usage

import asyncio

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

async def main() -> None:
    # Create a persistent session first — the toolset never creates one
    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("Plot a histogram of [1, 2, 2, 3, 3, 3, 4] with matplotlib")
        print(result.output)
    finally:
        await toolset.aclose()
        await session.destroy()  # this app owns the session

asyncio.run(main())

What inis_toolset() provides

The toolset holds a persistent session for the whole conversation — multiple tool calls, and multiple agent.run() calls, share state. session_id is required; the toolset only ever attaches, never creates or destroys a session.

This is deliberately the smallest execution/file surface. Session administration (pause, resume, checkpoint, restore, expose_port, capture_artifacts) and unbounded stateless-execution sandbox creation (execute_once) are opt-in from inis_pydantic_ai.lifecycle — add them to toolsets=[...] only if your app wants the model to have that control.

For a capability-based alternative that follows the current pydantic-ai-harness sandbox contract (run_command, read_file, write_file, list_directory), see InisSandbox in inis_pydantic_ai.sandbox — same attach-only session rule, registered via Agent(capabilities=[...]) instead of toolsets=[...].

Configuration

inis_toolset(
    session_id="9hmquuxiz7qyqdqiwfvep",  # required
    api_key="inis_...",           # or $INIS_API_KEY
    base_url="https://api.inis.run",
)

vs LangChain

Both wrap the same Python SDK. Pick one — or use remote MCP and skip framework glue entirely.