kloakd-sdk — Official Python SDK for KLOAKD

The KLOAKD SDK gives Python developers typed access to all 7 KLOAKD modules. Modules compose like Unix pipes — the artifact output of one becomes the input to the next, eliminating redundant fetches.

Installation

pip install kloakd-sdk

Requires Python 3.9+ and httpx>=0.27 (installed automatically).

Quickstart

from kloakd import Kloakd

client = Kloakd(
    api_key="sk-live-...",
    organization_id="your-org-id",
)

# Step 1: Fetch with anti-bot bypass
fetch = client.evadr.fetch("https://books.toscrape.com")
print(f"Fetched via tier {fetch.tier_used}")

# Step 2: Crawl the site hierarchy (waits until the crawl finishes)
crawl = client.webgrph.crawl("https://books.toscrape.com", max_depth=2)
print(f"Found {crawl.total_pages} pages")

# Step 3: Extract structured data — reuse the Evadr artifact again
data = client.kolektr.page(
    "https://books.toscrape.com",
    schema={"title": "css:h3 a", "price": "css:p.price_color"},
    fetch_artifact_id=fetch.artifact_id,
)
for book in data.records[:3]:
    print(book)

The 7 Modules

Artifact Chaining

Artifacts are the composition primitive — the output of one module becomes the input to another:

evadr.fetch()    → artifact_id (FETCHED_CONTENT)
  └─→ kolektr.page(fetch_artifact_id=...)       # skip re-fetch

fetchyr.authenticate() → artifact_id (AUTHENTICATED_SESSION)
  └─→ evadr.fetch(session_artifact_id=...)      # authenticated fetch
  └─→ webgrph.crawl(session_artifact_id=...)    # authenticated crawl
  └─→ kolektr.page(session_artifact_id=...)

webgrph.crawl()  → crawl_id (SITE_HIERARCHY)
  └─→ skanyr.site_from_crawl(crawl_id)          # into the site store

Async Client

from kloakd import AsyncKloakd

client = AsyncKloakd(api_key="sk-live-...", organization_id="your-org-id")

async def main():
    fetch = await client.evadr.fetch("https://books.toscrape.com")
    data  = await client.kolektr.page("https://books.toscrape.com",
                                       fetch_artifact_id=fetch.artifact_id)

    # SSE stream: pipeline events (type is pipeline_step, pipeline_error
    # or pipeline_complete)
    async for event in client.webgrph.crawl_stream("https://books.toscrape.com"):
        print(event.type, event.stage_id, event.message, event.progress_pct)

The sync and async clients have the same methods with the same arguments.

Error Handling

from kloakd.errors import KloakdError, RateLimitError, AuthenticationError
import time

try:
    result = client.evadr.fetch("https://example.com")
except RateLimitError as e:
    time.sleep(e.retry_after)
except AuthenticationError:
    # Rotate API key
    raise
except KloakdError as e:
    print(f"Error {e.status_code}: {e.message}")

Error types:

Every error has message, status_code and code (the API's error code, such as rate_limited, or None). A response that is not JSON (a proxy's error page) does not crash the client: its text becomes the message.

Retries: 429 rate_limited and 503 billing_unavailable are retried on any method, after Retry-After (or until reset_at). Other 5xx answers and network errors are retried only for GET, HEAD and DELETE: a POST, PUT or PATCH may already have done its work, so it is never sent twice. At most max_retries attempts (default 3); timeout (default 100 s) applies to each attempt.

Long requests: the API answers within about 85 seconds (it sits behind Cloudflare, which closes requests at 100 s). A request still running then continues as a job: the API answers HTTP 202 with a poll_url. The SDK follows it and returns the same typed result as a direct answer. This applies to evadr.fetch, kolektr.page / page_all, fetchyr.authenticate, detect_forms, fill_form, detect_mfa and nexus.analyze, synthesize and execute. The first poll is after poll_interval seconds (default 3), each later wait half as long again up to 15 s, for at most poll_timeout seconds (default 900); set both on the client (Kloakd(..., poll_interval=, poll_timeout=)), or per call on evadr.fetch. A failed job, or one whose result has expired (an hour after it ends), raises JobFailedError; one still running at the timeout raises JobTimeoutError with its job_id. Resume it with client.wait_for_handoff(e.job_id) (or read its state once with client.get_handoff(job_id)); a fetch job resumes with client.evadr.wait_for_job(e.job_id). The default timeout of 100 s is above the API's 85 s wait, so the client receives the 202 rather than timing out first.

Nexus knowledge: nexus.execute(...) returns execution_result_id; pass it to nexus.knowledge(result.execution_result_id).

Crawls and pages

crawl = client.webgrph.crawl("https://example.com")   # waits, then loads every page
if not crawl.ok:
    print(crawl.status, crawl.error)
all_pages = crawl.pages
page = client.webgrph.get_crawl_pages(crawl.crawl_id, limit=100)   # one page, by cursor
all_records = client.kolektr.page_all("https://example.com")

# Discover, fetch and extract in one call. A page that fails is marked
# failed; billing, plan, key and rate-limit errors are raised.
site = client.crawl("https://example.com", max_depth=2,
                    extract_schema={"title": "css:h1"})
print(site.success, site.error, site.pages_fetched, site.pages_failed)

# Store the crawl as a site, then search it
stored = client.skanyr.site_from_crawl(crawl.crawl_id)
hits = client.skanyr.search("pricing", site_id=stored["site_id"])
zip_bytes = client.skanyr.export_site(stored["site_id"])   # a ZIP archive

Extraction schemas are the shorthand {"field": "css:selector"} (the css: prefix is optional), a list of fields [{"name": "link", "selector": "a", "attribute": "href"}], or a full schema_spec with fields.

Parlyr — chat

turn = client.parlyr.chat("my-session", "Get book titles from https://books.toscrape.com")
print(turn.response)          # text and clarification content
print(turn.interpretation)    # the plan: url, entity, fields, summary, ...
print(turn.results)           # data of each results event
if not turn.ok:
    print([e.content for e in turn.errors])

for event in client.parlyr.chat_stream("my-session", "Also get prices"):
    print(event.type, event.content, event.data, event.actions)

Fetchyr — RPA & Authentication

# Store credentials once, then log in and produce a reusable session
client.fetchyr.store_credentials("app", username="user@example.com", password="secret")
session = client.fetchyr.authenticate("app", "https://app.example.com/login")

# Detect and handle MFA
mfa = client.fetchyr.detect_mfa("https://app.example.com/mfa",
                                  session_artifact_id=session.artifact_id)
if mfa.mfa_detected:
    challenge = client.fetchyr.submit_mfa(mfa.challenge_id, "123456")

# Detect forms, run workflows, deduplicate records
forms  = client.fetchyr.detect_forms("https://example.com/signup")
seen   = client.fetchyr.check_duplicate("https://example.com/item/1")

Links

• Documentation: https://docs.kloakd.dev/sdk/python
• API Reference: https://docs.kloakd.dev/api
• Dashboard: https://app.kloakd.dev
• GitHub: https://github.com/kloakd-dev/kloakd-sdk-python
• PyPI: https://pypi.org/project/kloakd-sdk/

License

MIT