smolagents

smolagents (Hugging Face) executes tools through Tool.__call__, which funnels into the tool's forward method. The AxioRank adapter shadows forward on the instance with a guarded version: the tool's name, description, and input schema are untouched, so the guarded tool is a drop-in replacement in any CodeAgent or ToolCallingAgent toolbox.

Install

pip install axiorank smolagents

The adapter is framework-free (it only wraps the instance method), so it works with any smolagents version that has Tool.forward.

Guard a toolbox

from smolagents import CodeAgent, InferenceClientModel, WebSearchTool
from axiorank import AxioRank
from axiorank.integrations.smolagents import guard_tools

axio = AxioRank(api_key="axr_live_...")

agent = CodeAgent(
    tools=guard_tools([WebSearchTool()], axio, on_deny="return"),
    model=InferenceClientModel(),
)
agent.run("Find the latest MCP security advisories.")

Every tool execution is scored first. On a deny, on_deny="return" hands the agent a short refusal string it can recover from; the default on_deny="raise" raises AxioRankDeniedError and fails the step hard.

Correlate a run

Pass a trace handle instead of the bare client so the gateway can correlate the whole run into a kill-chain trace:

with axio.trace() as t:
    agent = CodeAgent(tools=guard_tools(tools, t), model=model)
    agent.run(task)

Output inspection

Pass inspect_results=True to also score each untrusted-source tool's output (web results, fetched pages) for indirect prompt injection before the model ingests it. See Tool-output inspection.