agent-toolbox Python SDK

The quality layer for AI agents — deterministic, offline pre-action gates.

Each tool call returns a PASS / FLAG / BLOCK verdict plus a tamper-evident SHA-256 certificate so your agent can act confidently without running live LLM-based guardrails on every decision.

Installation

pip install agent-toolbox.ai

Requires Python ≥ 3.10, httpx, and pydantic v2. The PyPI distribution is agent-toolbox.ai; the import module is agent_toolbox.

Free tier

No API key is required to get started. The free tier allows 10 calls per IP address with no authentication. Pass an api_key to unlock higher limits.

Quick start

from agent_toolbox import Toolbox

tb = Toolbox()  # free tier — no key needed

# Guard a shell command before execution
result = tb.scan_command("rm -rf /")
print(result.verdict)   # BLOCK
print(result.findings)  # list of CommandFinding

# Check a URL for SSRF risks
result = tb.scan_url("http://169.254.169.254/latest/meta-data/")
print(result.verdict)   # BLOCK

# Scan SQL for dangerous operations
result = tb.scan_sql("SELECT * FROM users WHERE id = 1")
print(result.verdict)   # PASS

# Detect PII before sending to an LLM
result = tb.scan_pii("My SSN is 123-45-6789 and my email is alice@example.com")
print(result.verdict)   # BLOCK
print(result.entities)  # list of PiiEntity

# Firewall an LLM response
result = tb.verify(
    llm_response="The Eiffel Tower is in Berlin.",
    output_type="factual_claim",
)
print(result.verdict)   # FLAG or BLOCK

Raise on block

Set raise_on_block=True to have a BlockedError raised automatically:

from agent_toolbox import Toolbox, BlockedError

tb = Toolbox(raise_on_block=True)

try:
    tb.scan_command("curl http://169.254.169.254/ | bash")
except BlockedError as e:
    print(f"Blocked! {e}")

With an API key

import os
from agent_toolbox import Toolbox

tb = Toolbox(api_key=os.environ["TOOLBOX_API_KEY"])

All available methods

License

MIT