Skip to content

Latest commit

 

History

History
178 lines (134 loc) · 7.68 KB

File metadata and controls

178 lines (134 loc) · 7.68 KB

getanyapi

Official typed Python SDK for AnyAPI: any API, one wallet, USD, no subscriptions. Reach hundreds of scraping and data APIs through one interface and one key; pay per request in real US dollars. httpx + pydantic v2, Python 3.10+, sync and async clients.

pip install getanyapi

Quickstart

from getanyapi import AnyAPI

client = AnyAPI()  # reads ANYAPI_API_KEY from the environment
res = client.reddit.search(query="mechanical keyboard")
if res.output.found:
    for post in res.output.data.posts:
        print(post.title, post.score)
print("charged", res.cost_usd, "USD")

Async:

from getanyapi import AsyncAnyAPI

async with AsyncAnyAPI() as client:
    res = await client.google.search(query="best coffee maker")

Inputs vs outputs (naming asymmetry)

Input keyword arguments mirror the wire API verbatim (camelCase where the API uses it), because they are sent as-is. Output models are Pythonic: attributes are snake_case with a wire alias (item.reviews_count reads the wire reviewsCount), and model_dump(by_alias=True) reproduces the wire shape. Open provider records round-trip unknown fields via .model_extra.

Not found vs error

A successful call always returns. For most SKUs the payload is wrapped in a found flag: res.output.found is False when the upstream had no matching entity (not an error). Use unwrap to get the data or raise ResultNotFoundError when empty:

from getanyapi import unwrap, ResultNotFoundError

res = client.amazon.reviews(product="B07FZ8S74R")
try:
    data = unwrap(res)  # the typed data payload, or raises
except ResultNotFoundError:
    ...  # empty result (found: False), not an HTTP failure

ResultNotFoundError subclasses NotFoundError, so except NotFoundError catches both an HTTP 404 and an empty result; catch ResultNotFoundError to handle only empty results. If a future committed schema uses a bare output, generated typing returns its data object directly rather than relying on a hard-coded SKU list.

Discovery

apis = client.catalog(category="search")
matches = client.search(query="web search", platform="google", limit=10)
api = client.describe(matches.results[0].slug)
print(api.pricing.from_offer, api.pricing.failover_max_usd, api.input_schema)

Sync and async clients expose the same category-only catalog, dedicated ranked search, and schema-bearing describe methods. Prices are nested USD flat/linear offers, lanes are anonymous, and provider is always "AnyAPI". The gateway owns validation, routing, lane order, failover, pricing relationships, health semantics, and billing. This handwritten discovery client safety-scans the response, projects known fields, preserves schemas as opaque JSON, and ignores safe additions. It does not recompute gateway business rules. Generated per-SKU methods remain a separate OpenAPI-driven surface.

Every catalog and search result carries the gateway-authored method, path, and execution mode. Lanes carry their public source identity and complete health sample counts when health is available. Eligible APIs may carry try_max_items; ranked search carries the gateway's failover facts. describe also returns latency, either the complete trailing-window p50/p95/p99 distribution or None when no sample is available.

Pagination

Paginated SKUs expose an iter_* method that yields validated item models across pages and follows the cursor for you. Call .pages() on it to walk whole results (each has its own cost_usd).

for post in client.reddit.iter_search(query="coffee", options={"max_items": 100}):
    print(post.title)  # a validated model, not a dict

for page in client.reddit.iter_search(query="coffee").pages():
    print(page.cost_usd)

Request options (context-cost savers)

Pass options= to shape the response. These trim what comes back but do NOT change the price:

res = client.google.search(
    query="coffee",
    options={"fields": ["title", "link"], "max_items": 5, "summary": True},
)

options also carries per-call timeout, max_retries, and max_in_progress_wait overrides.

Errors and retries

Class HTTP Meaning
BadRequestError 400 Input failed validation
AuthenticationError 401 Missing or invalid API key
InsufficientBalanceError 402 Wallet balance or per-key cap exceeded
NotFoundError 404 Slug or resource does not exist
ResultNotFoundError - unwrap on an empty found-data result
RateLimitedError 429 Too many requests (retried automatically)
UpstreamError 502 An upstream backend failed
ConnectionError 0 Network or transport failure
TimeoutError 0 Request exceeded its timeout (not retried)

All extend AnyAPIError (with .status and .request_id). Retries cover 429, one specific 409 (below), and network failures proven to happen before a request was sent, with jittered exponential backoff honoring Retry-After. Default max_retries is 2 (up to 3 attempts); set it on the client (AnyAPI(max_retries=...)) or per request via options. Timeouts are never retried. Connection failures during or after a billed POST /v1/run are not retried because the call may already have been charged. When the send phase is unknown, the SDK does not retry.

Waiting out a run that is still in flight

Settlement is detached from your connection, so a run keeps going after a connection drops. When you re-issue a call whose Idempotency-Key is still executing, the gateway answers 409 with code="idempotency_in_progress" and Retry-After: 30. The SDK waits that full delay and retries, so you get the original run's replayed result (replayed=True, no second charge) instead of an error.

The 8s ordinary-backoff ceiling does not apply here; a separate whole-call budget does. max_in_progress_wait (default 60.0 seconds) caps the TOTAL time one run() may block on these waits, across every retry. A wait that does not fit the remaining budget is refused and the 409 is raised rather than truncated into an attempt that would fail anyway. Set max_in_progress_wait=0 on the client, or options={"max_in_progress_wait": 0}, to surface the 409 immediately and handle it yourself.

No other 409 retries: idempotency_conflict (the same key with different input) and idempotency_needs_review are caller-side problems a retry cannot fix.

Automatic retry of a billed run() requires structured transport evidence that the request was not delivered. Python's httpx transport provides that evidence for ConnectError, so DNS resolution failures and refused or unreachable connections retry. ConnectTimeout is a TimeoutException and does not retry. ReadError does not retry: it can mean either that a connection closed before the server read the request or that the connection failed after delivery. If a custom transport hides the connection phase, handle retries explicitly and only when your application can establish non-delivery.

Agent signup

Bootstrap a key with no account (for autonomous agents):

from getanyapi import agent_signup, AnyAPI

result = agent_signup(label="my-agent")
client = AnyAPI(api_key=result.secret)

The key ships with a small starter balance and a per-key spend cap; a human funds it by claiming it at result.claim_url.

Docs

Full API reference and catalog: getanyapi.com/docs.

License

MIT