driftgard

Official Python SDK for DriftGard.

Use it to control what production AI can say and do: evaluate prompts and responses, run OpenAI-compatible gateway calls through DriftGard, validate agent tool calls, and record audit evidence.

Install

pip install driftgard

Requires Python 3.8 or newer.

Quick Start

from driftgard import Driftgard

dg = Driftgard(api_key="dg_...")

result = dg.evaluate(
    project_id="proj_...",
    prompt="What stocks should I buy?",
    response="You should invest everything in...",
    model_id="gpt-4o",
)

if not result["evaluation"]["allowed"]:
    print("Blocked:", result.get("fallback", {}).get("message"))
elif result["evaluation"].get("sanitized"):
    print("Use sanitized response:", result["evaluation"]["sanitized_response"])
else:
    print("Allowed")

Inline Gateway

Use the gateway when DriftGard should sit in the request path, call the provider, enforce policy, and return the final response.

completion = dg.gateway_chat_completions(
    project_id="proj_...",
    model="gpt-4o-mini",
    messages=[
        {"role": "system", "content": "You are a support assistant."},
        {"role": "user", "content": "Summarize the refund process."},
    ],
    driftgard={
        "gateway_mode": "enforce",
        "agent_id": "support-agent",
        "agent_role": "customer_support",
        "session_id": "sess_123",
    },
)

print(completion["choices"][0]["message"]["content"])
print(completion.get("driftgard", {}).get("gateway_decision"))

Streaming is also supported:

stream = dg.stream_gateway_chat_completions(
    project_id="proj_...",
    model="gpt-4o-mini",
    messages=messages,
    driftgard={"gateway_mode": "observe"},
)

Tool Call Validation

Validate agent tool calls against Control Pack tool_rules.

result = dg.evaluate(
    project_id="proj_...",
    eval_mode="tool_call",
    tool_call={
        "tool_name": "transfer_money",
        "parameters": {"amount": 500, "to_account": "acct_123"},
    },
    agent_role="payments_agent",
    model_id="agent-runtime",
)

if not result["evaluation"]["allowed"]:
    raise RuntimeError(result.get("fallback", {}).get("message", "Tool call blocked"))

Local Evaluation

For sensitive environments, the SDK can evaluate locally through DriftGard's WASM evaluator.

dg = Driftgard(
    api_key="dg_...",
    mode="local",
    project_id="proj_...",
)

dg.init()
result = dg.evaluate(prompt=prompt, response=response, model_id="local-model")

Modes:

Mode	Behavior
`remote`	Send prompt/response to DriftGard API
`local`	Evaluate locally; no prompt/response leaves your environment
`local-with-audit`	Evaluate locally and report verdict metadata only

Framework Integrations

Included integrations:

LangChain response guardrail
LangChain tool guard
CrewAI response and tool guardrails
Strands / AWS Agent Core hook
local semantic matching for local mode

Documentation

Requirements

Python 3.8+
DriftGard API key
DriftGard project ID

License

MIT

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

driftgard-1.29.0.tar.gz (500.1 kB view details)

Uploaded Aug 12, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

driftgard-1.29.0-py3-none-any.whl (490.9 kB view details)

Uploaded Aug 12, 2026 Python 3

File details

Details for the file driftgard-1.29.0.tar.gz.

File metadata

Download URL: driftgard-1.29.0.tar.gz
Upload date: Aug 12, 2026
Size: 500.1 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/6.2.0 CPython/3.9.6

File hashes

Hashes for driftgard-1.29.0.tar.gz
Algorithm	Hash digest
SHA256	`639e2d2d38410adb82a2ccfd654f959a0883d7bf61cda61ef1d44cf91dfa1556`
MD5	`29616525e05a31eebff2cd5576e2f028`
BLAKE2b-256	`30caed4288d8dfe927a27ace83acd72f9e366f18edf2c396d09453a63d3ac847`

See more details on using hashes here.

File details

Details for the file driftgard-1.29.0-py3-none-any.whl.

File metadata

Download URL: driftgard-1.29.0-py3-none-any.whl
Upload date: Aug 12, 2026
Size: 490.9 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/6.2.0 CPython/3.9.6

File hashes

Hashes for driftgard-1.29.0-py3-none-any.whl
Algorithm	Hash digest
SHA256	`a00e56ae4dcb2ce9b4a05b86487171a690a83472a9f169ada1789926532c7514`
MD5	`c31d4c725305b39fe4011068a452c133`
BLAKE2b-256	`36a5c63b258d4cac7e0f2ae45f9679b2e5eb50a2cb29fb118f25a70842bd5341`