gatewerk 0.1.1

Python SDK for Gatewerk — human oversight for AI agents

gatewerk

Python SDK for Gatewerk — the open-source human oversight station for AI agents.

Install

pip install gatewerk

Quick Start

from gatewerk import create_client

gw = create_client(
    api_key="gw_key_...",
    url="https://api.gatewerk.com",
)

# Submit a review request
review = gw.reviews.create(
    "email-review",
    payload={"to": "[e-mail]", "subject": "Q4 Report", "body": draft},
    callback_url="https://my-agent.com/webhook",
    priority="high",
)
print(review.id)      # gw_rev_...
print(review.status)  # pending

Async Support

from gatewerk import create_async_client

async with create_async_client(api_key="gw_key_...") as gw:
    review = await gw.reviews.create("email-review", payload={...})
    feedback = await gw.feedback.query(template="email-review")

Framework integrations

Framework	Install	Status
LangGraph	pip install gatewerk[langgraph]	available
CrewAI	pip install gatewerk[crewai]	available

LangGraph in 30 seconds:

from gatewerk import create_client
from gatewerk.integrations.langgraph import gatewerk_interrupt

gw = create_client(api_key="gw_key_...")

def gated_node(state):
    decision = gatewerk_interrupt(gw, template="refund_approval", payload={...})
    return {"approved": decision.approved}

See gatewerk/integrations/langgraph/README.md for the full guide, including the polling alternative (await_decision), async variants, and the LangGraph re-execution caveat.

CrewAI in 30 seconds:

from crewai import Agent, Task, Crew
from gatewerk import create_client
from gatewerk.integrations.crewai import GatewerkApprovalTool

gw = create_client(api_key="gw_key_...")
approval_tool = GatewerkApprovalTool(gw, template="refund_approval")

agent = Agent(role="Customer Support", goal="...", tools=[approval_tool])
Crew(agents=[agent], tasks=[Task(description="...", agent=agent, expected_output="...")]).kickoff()

The tool pauses the agent, routes the request to the Gatewerk dashboard, and returns "APPROVED", "REJECTED", or "CHANGES_REQUESTED: ..." once a human decides. See gatewerk/integrations/crewai/README.md for the full guide.

Resources

Resource	Methods
gw.reviews	create() get() list() decide() retry() update() cancel_request() versions() create_token() list_auto_paginate()
gw.chains	create() get() get_for_review()
gw.notes	create() get() list() list_auto_paginate() update() delete() pin() unpin() tags()
gw.templates	list() get() create() update() delete() stats()
gw.feedback	query()
gw.audit	query()
gw.stats	get()
gw.webhooks	verify() deliveries()
gw.key_info()	Introspect API key scopes

Typed Responses

All methods return Pydantic models:

from gatewerk import Review, ReviewList, Template, FeedbackList

review = gw.reviews.get("gw_rev_...")
review.id              # str
review.status          # str
review.payload         # dict[str, Any]
review.current_version # int
review.auto_approved   # bool | None

Auto-Pagination

# Iterate all reviews without managing offsets
for review in gw.reviews.list_auto_paginate(status="pending"):
    print(review.id)

# Async
async for review in gw.reviews.list_auto_paginate(status="pending"):
    print(review.id)

Retry & Resilience

The SDK automatically retries on transient errors (429, 500, 502, 503, 504) with exponential backoff and jitter. Rate limit Retry-After headers are respected.

# Customize retry behavior
gw = create_client(
    api_key="gw_key_...",
    max_retries=5,   # default: 2, set 0 to disable
    timeout=60.0,    # default: 30s
)

Error Handling

from gatewerk import (
    GatewerkError,        # Base — all errors
    InvalidRequestError,  # 400
    AuthenticationError,  # 401
    ForbiddenError,       # 403
    NotFoundError,        # 404
    ConflictError,        # 409
    RateLimitError,       # 429 (includes .retry_after)
)

try:
    review = gw.reviews.get("gw_rev_nonexistent")
except NotFoundError as e:
    print(e.message)      # "Review not found"
    print(e.status_code)  # 404
    print(e.code)         # "review_not_found"
except RateLimitError as e:
    print(e.retry_after)  # seconds until retry

Webhook Verification

Every delivery carries two signature headers. Verify either depending on whether you need replay protection.

v1 (authenticity only, no replay protection)

The SDK's gw.webhooks.verify() helper accepts the v1 X-Webhook-Signature header:

# In your Flask/FastAPI webhook handler
payload = gw.webhooks.verify(
    raw_body=request.body,
    signature_header=request.headers["X-Webhook-Signature"],
    secret="whsec_...",
)
# payload is a trusted dict: {"event": "review.decided", ...}

v2 (replay-safe; recommended for new integrations)

The X-Webhook-Signature-V2 header carries t=<unix-seconds>,v1=<hex> where hex = HMAC(f"{t}.{body}", secret). Parse, enforce freshness, then constant-time compare. No SDK helper yet — follow the manual pattern:

import hmac
import hashlib
import time

def verify_v2(raw_body: bytes, header: str, secret: str, tolerance_seconds: int = 300) -> bool:
    parts = dict(p.split("=", 1) for p in header.split(","))
    ts = int(parts["t"])
    received_hex = parts["v1"]

    # Freshness: reject replays outside the tolerance window.
    if abs(time.time() - ts) > tolerance_seconds:
        return False

    expected_hex = hmac.new(
        secret.encode(),
        f"{ts}.{raw_body.decode()}".encode(),
        hashlib.sha256,
    ).hexdigest()
    return hmac.compare_digest(expected_hex, received_hex)

assert verify_v2(
    request.body,
    request.headers["X-Webhook-Signature-V2"],
    "whsec_...",
)

Receivers should also dedup on X-Webhook-Id (stable across retries) to survive the legitimate retry flow. A v2-verified payload MAY still be a legitimate retry; the header id is the idempotency key.

Logging

The SDK logs via Python's standard logging module under the gatewerk logger:

import logging
logging.getLogger("gatewerk").setLevel(logging.DEBUG)

Environment Variables

Variable	Description
GATEWERK_API_KEY	API key (fallback if not passed to create_client)
GATEWERK_URL	API URL (default: http://localhost:3100)

License

Apache-2.0