snipget-client

snipget-client

The official Python client for Snipget, the hosted utility API for AI agents: data normalization, parsing, validation, and classification over plain HTTPS.

This SDK fronts the hosted Snipget MCP server (https://mcp.snipget.ai/mcp) — AI agents can connect directly, no install and no API key needed; see llms-install.md.

What is Snipget

Snipget is a hosted, pay-per-call utility API built for AI agents and the developers who build them. It serves 300+ programmatic endpoints for data normalization, parsing, validation, and classification, with particular depth in the life sciences: healthcare (NPI validation and lookup, DEA numbers, provider taxonomy, credentials, certifications), chemistry (compound lookup, CAS and SMILES validation, molecular weight, GHS hazard classification — backed by PubChem), and biotech (gene lookup and validation via HGNC, protein lookup via UniProt, drug normalization and synonyms via RxNorm, and clinical-trial lookup via ClinicalTrials.gov). Every endpoint is deterministic (no LLM calls inside the API), returns a confidence score, and ships in single-record and batch variants.

Snipget is agent-native by design. Agents can discover and call it through the OpenAPI spec or the MCP server, and every response uses one consistent JSON envelope so a single integration covers the whole catalog. This package is a thin HTTP wrapper around that hosted API; all the actual logic runs server-side, and the interactive docs are the per-endpoint contract.

Install

pip install snipget-client

Requires Python 3.10+. The only dependency is httpx.

Quickstart

You need an API key from snipget.ai. One generic call() method reaches every endpoint; pass the path and the JSON payload from the API docs.

from snipget import Client

client = Client(api_key="YOUR_API_KEY")  # or set SNIPGET_API_KEY

resp = client.call("/healthcare/npi/validate", {"npi": "1234567893"})

print(resp.result)
# {'npi': 1234567893, 'is_valid': True, 'checksum_valid': True, 'input_was_clean': True}
print(resp.confidence)        # 1.0
print(resp.meta.cost_units)   # 1
print(resp.meta.request_id)   # 'req_...'

# Batch variants exist for every utility:
resp = client.call(
    "/healthcare/npi/validate/batch",
    {"items": ["1234567893", "1234567890"]},
)
print(resp.result["summary"])  # {'total': 2, 'valid': 1, 'invalid': 1}

Async, same surface:

import asyncio
from snipget import AsyncClient

async def main():
    async with AsyncClient() as client:  # reads SNIPGET_API_KEY
        resp = await client.call(
            "/common/phone/validate",
            {"value": "(415) 555-0132", "country_hint": "US"},
        )
        print(resp.result)

asyncio.run(main())

call() defaults to POST when a payload is given and GET otherwise, which matches every endpoint in the spec; pass method= to override.

Authentication

Get an API key at snipget.ai. The client resolves the key in this order:

1. Client(api_key="...")
2. The SNIPGET_API_KEY environment variable

By default the key is sent as Authorization: Bearer <key>. The API also accepts an X-API-Key header; opt in with Client(auth_header="x-api-key").

Error handling

Every API error is raised as a typed exception. All of them subclass SnipgetError and carry error_code, message, request_id, http_status, and the full parsed envelope as body.

import snipget

client = snipget.Client()

try:
    resp = client.call("/healthcare/npi/validate", {"npi": "1234567893"})
except snipget.AuthenticationError as e:
    print("Check your API key:", e.error_code)            # 401/403
except snipget.InvalidRequestError as e:
    print("Bad request:", e.body.get("details"))           # 400/422
except snipget.RateLimitError as e:
    print("Throttled; retry in", e.retry_after, "seconds")  # 429 RATE_LIMITED
except snipget.QuotaExceededError as e:
    print("Out of monthly capacity:", e.body.get("limit_type"))
    print("Allowance left (USD):", e.credit_remaining_usd)  # 429 QUOTA_EXCEEDED
except snipget.MaintenanceError as e:
    print("Maintenance window; retry in", e.retry_after)    # 503 MAINTENANCE_MODE
except snipget.UpstreamRateLimitedError as e:
    print("Upstream throttled us; retry in", e.retry_after)  # 503 UPSTREAM_RATE_LIMITED
except snipget.UpstreamError as e:
    print("Upstream data source is down:", e.error_code)     # 503 UPSTREAM_UNAVAILABLE
except snipget.APIError as e:
    print("Server error; quote this id to support:", e.request_id)

The two 429s mean different things: RateLimitError is a per-second throughput throttle and clears in seconds; QuotaExceededError means the monthly included calls or prepaid overage allowance are exhausted and will not clear until the monthly reset, a tier upgrade, or an allowance top-up. The client retries the first automatically and never retries the second.

A few utilities call external data sources (PubChem, RxNorm, ClinicalTrials.gov, the FX feed, NPPES). When one of those is down you get an UpstreamError (503 UPSTREAM_UNAVAILABLE); when one is throttling Snipget you get an UpstreamRateLimitedError (503 UPSTREAM_RATE_LIMITED, carrying retry_after) — distinct from RateLimitError, because your own request rate is fine. Both subclass APIError, so a single except snipget.APIError still catches them; both are transient and retried automatically.

Retries and timeouts

client = Client(
    api_key="...",
    timeout=30.0,      # per-request timeout in seconds
    max_retries=2,     # retries on top of the initial attempt
)

The client automatically retries network errors, RATE_LIMITED and UPSTREAM_RATE_LIMITED (honoring the server's Retry-After), and 5xx responses (including UPSTREAM_UNAVAILABLE), using exponential backoff with jitter. Snipget utility calls are pure and idempotent, so retrying a POST is safe. It never retries QUOTA_EXCEEDED or any other 4xx. Maintenance 503s are retried on the short backoff only; if the window outlasts the retry budget you get a MaintenanceError with retry_after (typically 300 seconds) so you can schedule your own retry.

The response envelope

Every Snipget endpoint, success or error, returns one envelope shape. call() returns a SnipgetResponse:

Meta fields the server didn't send are None; unknown future fields stay available via meta.raw.

Links

• Website: https://snipget.ai
• Interactive API docs: https://api.snipget.ai/docs
• OpenAPI spec: https://api.snipget.ai/openapi.json
• npm sibling: snipget on npm — the official JavaScript/TypeScript client

License

MIT. Copyright 2026 Snipget Inc.