superagent-ai 0.0.15

Python SDK for Superagent.

Superagent Python SDK

Python client for calling the Superagent Guard and Redact endpoints.

Installation

pip install superagent-ai

Local development with uv

From the repository root, install the package (including test extras) and create a managed virtual environment:

cd sdk/python
uv sync --extra tests

This will provision .venv, install the SDK in editable mode, and pull in the testing dependencies. Once synced, run the test suite with:

uv run pytest tests

Usage

import asyncio
from superagent_ai import create_client

async def main() -> None:
    client = create_client(
        api_base_url="https://app.superagent.sh/api",  # Optional, this is the default
        api_key="sk-...",
    )

    # Guard: Analyze commands for security threats
    guard_result = await client.guard(
        "Write a hello world script",
        on_block=lambda reason: print("Guard blocked:", reason),
        on_pass=lambda: print("Guard approved!"),
    )

    if guard_result.rejected:
        print("Rejected:", guard_result.reasoning)
    else:
        print("Approved:", guard_result.decision)

    # Redact: Remove sensitive data from text
    redact_result = await client.redact(
        "My email is [e-mail] and SSN is 123-45-6789"
    )

    print(redact_result.redacted)
    # Output: "My email is <REDACTED_EMAIL> and SSN is <REDACTED_SSN>"

    await client.aclose()

asyncio.run(main())

Using as a context manager

import asyncio
from superagent_ai import create_client

async def main() -> None:
    async with create_client(api_key="sk-...") as client:
        result = await client.guard("command")
        redacted = await client.redact("text")

asyncio.run(main())

API Reference

create_client(**kwargs)

Creates a new Superagent client.

Parameters:

• api_key (required) – API key provisioned in Superagent
• api_base_url (optional) – Base URL for the API (defaults to https://app.superagent.sh/api)
• client (optional) – Custom httpx.AsyncClient instance
• timeout (optional) – Request timeout in seconds (defaults to 10.0)

Returns: Client

client.guard(command, *, on_block=None, on_pass=None)

Analyzes a command for security threats.

Parameters:

• command – The text to analyze
• on_block (optional) – Callback function called when command is blocked
• on_pass (optional) – Callback function called when command is approved

Returns: GuardResult

@dataclass
class GuardResult:
    rejected: bool              # True if guard blocked the command
    reasoning: str              # Explanation from the guard
    raw: AnalysisResponse       # Full API response
    decision: Optional[GuardDecision]  # Parsed decision details
    usage: Optional[GuardUsage]        # Token usage statistics

@dataclass
class GuardDecision:
    status: Literal["pass", "block"]
    violation_types: list[str]
    cwe_codes: list[str]

client.redact(text, *, url_whitelist=None, entities=None, file=None, format=None)

Redacts sensitive data from text.

Parameters:

• text – The text to redact
• url_whitelist (optional) – List of URL prefixes that should not be redacted
• entities (optional) – List of custom entity types to redact (natural language descriptions)
• file (optional) – File object to redact (e.g., PDF document)
• format (optional) – Format of the file (currently only "PDF" is supported)

Returns: RedactResult

@dataclass
class RedactResult:
    redacted: str               # Text with sensitive data redacted
    reasoning: str              # Explanation of what was redacted
    raw: dict                   # Full API response
    usage: Optional[GuardUsage] # Token usage statistics

Detected PII/PHI Types

The redaction feature detects and replaces:

• Email addresses → <REDACTED_EMAIL>
• Social Security Numbers → <REDACTED_SSN>
• Credit cards (Visa, Mastercard, Amex) → <REDACTED_CC>
• Phone numbers (US format) → <REDACTED_PHONE>
• IP addresses (IPv4/IPv6) → <REDACTED_IP>
• API keys & tokens → <REDACTED_API_KEY>
• AWS access keys → <REDACTED_AWS_KEY>
• Bearer tokens → Bearer <REDACTED_TOKEN>
• MAC addresses → <REDACTED_MAC>
• Medical record numbers → <REDACTED_MRN>
• Passport numbers → <REDACTED_PASSPORT>
• IBAN → <REDACTED_IBAN>
• ZIP codes → <REDACTED_ZIP>

Custom Entity Redaction

You can specify custom PII entities to redact using natural language:

client = create_client(api_key="sk-...")

result = await client.redact(
    "My credit card is 4532-1234-5678-9010 and employee ID is EMP-12345",
    entities=["credit card numbers", "employee IDs"]
)
# Output: "My credit card is <REDACTED> and employee ID is <REDACTED>"

URL Whitelisting

You can specify URLs that should not be redacted by passing the url_whitelist parameter:

client = create_client(api_key="sk-...")

result = await client.redact(
    "Check out https://github.com/user/repo and https://secret.com/data",
    url_whitelist=["https://github.com", "https://example.com"]
)
# Output: "Check out https://github.com/user/repo and <URL_REDACTED>"

The whitelist is applied locally after redaction - URLs matching the prefixes are preserved, while non-whitelisted URLs are replaced with <URL_REDACTED>.

PDF File Redaction

You can redact sensitive information from PDF files:

import asyncio
from superagent_ai import create_client

async def main() -> None:
    async with create_client(api_key="sk-...") as client:
        # Read and redact PDF file
        with open("sensitive-document.pdf", "rb") as pdf_file:
            result = await client.redact(
                text="Analyze and redact PII from this document",
                file=pdf_file,
                format="PDF",
                entities=["SSN", "credit card numbers", "email addresses"]
            )

            print(result.redacted)  # Redacted text content from the PDF
            print(result.reasoning) # Explanation of what was redacted

asyncio.run(main())

Note: File redaction uses multipart/form-data encoding and currently supports PDF format only.

Error Handling

from superagent_ai import GuardError

try:
    result = await client.guard("command")
except GuardError as error:
    print(f"Guard error: {error}")