pydantic-ai-withruntime

Runtime sandboxes as a Pydantic AI toolset. Each agent run gets its own Runtime sandbox, a Firecracker microVM with its own Linux kernel. The sandbox is created when the run starts and stopped when it ends.

pip install pydantic-ai pydantic-ai-withruntime   # pydantic-ai brings the model providers
npx withruntime login   # or set RUNTIME_API_KEY

from pydantic_ai import Agent
from pydantic_ai_withruntime import RuntimeToolset

agent = Agent("anthropic:claude-sonnet-4-6", toolsets=[RuntimeToolset()])
print(agent.run_sync("Fit a line to (1,2), (2,4.1), (3,6.2) with numpy and give the slope.").output)

The model sees four tools:

Relative paths are under /workspace. Output past 20,000 characters is cut from the front, keeping the end, where errors are.

Options

• RuntimeToolset(create={...}) sets each run's sandbox: vcpu, memory_mib, image, network, labels and the rest of sandboxes.create. Unless you set a timeout_seconds, each sandbox gets a 30-minute lease that ends in a stop, so a crashed process leaves nothing running.
• RuntimeToolset(sandbox) uses an AsyncSandbox you created. Every run shares it, and it is never stopped for you. Use this to keep files between runs, for example with AsyncRuntime().sandboxes.get_or_create("user-42").
• timeout_seconds, max_output_chars and root tune the tools. instructions=None leaves the agent's instructions unchanged.

Tests

tests/ runs the toolset under Pydantic AI's own agent loop, with a scripted FunctionModel and a fake client; uv run --resolution lowest-direct --group test pytest runs them at the lowest versions pyproject.toml allows. scripts/e2e.py runs the agent loop against the real API with RUNTIME_API_KEY set. It passed on 25 September 2026 with Pydantic AI 2.50.0: two runs, each in its own sandbox, both stopped.