addrly-api

addrly-api

The official Python client for the Addrly API: Belgian addresses (BOSA Best-Address) and Belgian companies (BCE/KBO, NBB annual accounts, Moniteur belge).

pip install addrly-api

from addrly_api import Addrly

client = Addrly()  # reads ADDRLY_API_KEY
# or Addrly("ak_...") / Addrly(api_key=key), from your own settings

for hit in client.autocomplete("rue de la loi", limit=5, type="address"):
    print(hit.score, hit.label)

address = client.normalize("16 rue de la loi 1000 brux")
if address and address.match_similarity >= 0.9:
    print(address.normalized.one_line())  # Rue de la Loi 16, 1000 Bruxelles

company = client.company("BE 0123.456.749")  # any format, checked locally
print(company.display_name, company.seat().one_line())

Requires Python 3.9+. The only dependency is httpx.

Get a key

Sign up at addrly.be, then create a key on the API keys page. Read it from the environment, never from source:

export ADDRLY_API_KEY=ak_...

Addrly() reads the environment variable. When your application already holds the key in its own settings, pass it to the constructor instead: Addrly("ak_...") or Addrly(api_key=key). Same for AsyncAddrly.

A key is server-to-server. Do not ship one to a browser or a mobile app.

What you can call

Know before you spend

account() is the only free call. It answers the plan, the feature slugs, the rate limit and what is left of both meters, and the server answers it even while billing is suspended, so it is also where you find out why everything else is refusing you:

account = client.account()

if account.frozen:
    raise SystemExit("billing suspended, resolve it in the app")
if not account.has_feature("company_api_bulk"):
    raise SystemExit("this plan does not include the bulk matcher")
if account.requests_remaining < len(rows):
    rows = rows[: int(account.requests_remaining)]

print(account.plan.name, account.credits_remaining, "dossiers left")
print("watching", account.watch.used, "of", account.watch.limit, "companies")
print("resets in", account.period.resets_in_seconds, "s")

A budget guard never refuses it either: a status call you cannot make when you are out of everything is a status call for nothing.

Two meters, and they are separate: requests (every call that does not open a dossier, company identity reads and matched rows included) and dossiers (one company's full dossier, charged once per company per billing period). An answer carries only the meter it charged (X-RateLimit-* for a request, X-Credit-* for a dossier), so client.last_usage keeps the best known state of both, while response.usage stays strictly what that one response said:

client.normalize("rue de la loi 16")
print(client.last_usage.address.remaining)  # 993

client.dossier("0123456749")
print(client.last_usage.credits.remaining)  # 249: dossiers left
print(client.last_usage.address.remaining)  # 993, still known

answer = client.autocomplete("exampleco")
print(answer.usage.credits.is_empty)  # True: that call charged a request

Bulk: one bad row never loses the good ones

normalize_many, companies and dossiers chunk your list to the API's ceilings (100 rows, 25 for full dossiers) and give back one Outcome per input, in input order. A malformed row is a 422 on that item, not on the call:

for outcome in client.normalize_many(rows):
    if outcome:
        save(outcome.value.normalized)
    else:
        log(outcome.index, outcome.error.code)

match_companies is the CRM/ERP import endpoint: messy names in, the best Belgian company match out, with the scored runners-up.

for row in client.match_companies(
    ["acme sofware", {"name": "globex", "address": "1000 Bruxelles"}]
):
    if row and row.score >= 0.9:
        apply(row.enterprise_number)

Errors: branch on the code, never on the status

Every refusal is a typed exception carrying the API's own error code, the message, the request_id and whatever fields that code documents.

from addrly_api import CreditsExhausted, FeatureNotInPlan, RateLimited, APIError

try:
    dossier = client.dossier("0123456749")
except CreditsExhausted as exc:
    print("out of credits, resets in", exc.reset_in_seconds, "s")
except FeatureNotInPlan as exc:
    print("the plan does not include", exc.feature)
except APIError as exc:
    print(exc.code, exc.request_id)

Retries: what is repeated, and what must not be

Two 429s mean opposite things, so the client treats them differently.

• Retried, with exponential backoff plus jitter, honouring Retry-After: rate_limited, 5xx, and connection failures.
• Never retried: quota_exceeded and credits_exhausted. A monthly allowance does not come back in 500 ms, and a loop on it only buys overage. Also never retried: 401, 402, 403, 404, 422. The request is wrong, or the plan is, and sending it again changes neither.
• A priced POST (/normalize, /batch, /enrich/company) is metered upfront, so it is only retried when the connection provably never opened. A read timeout there might already have been billed, and the API has no idempotency key yet.

client = Addrly(max_retries=3, retry_backoff=0.5, timeout=10.0)

max_retries=0 disables it.

Never spend more than you meant to

A migration script is the one place a paid API gets expensive by accident. budget() counts the declared price of each call and refuses the one that would cross the cap, before sending it:

from addrly_api import BudgetExceeded

try:
    with client.budget(credits=50) as spent:
        for number in numbers:
            enrich(client.dossier(number))
except BudgetExceeded:
    print("stopped at", spent.credits_spent, "credits")

Async

Same methods, same models, same retry policy:

import asyncio
from addrly_api import AsyncAddrly


async def main():
    async with AsyncAddrly() as client:
        answer = await client.autocomplete("rue de la loi")
        print(len(answer))


asyncio.run(main())

Command line

The package installs an addrly command (and addrly-api, same thing, for a machine where the short name is taken), so the API can be tried in thirty seconds:

addrly account
addrly autocomplete "rue de la loi"
addrly normalize "rue de la loi 16 brux"
addrly company 0123456749
addrly dossier 0123456749
addrly reverse 50.8467 4.3499
addrly match "acme sofware" "globex grup"

# The job most integrations write first: an address column, cleaned.
addrly clean clients.csv --column address -o clean.csv

clean adds addrly_street, addrly_house_number, addrly_box, addrly_postcode, addrly_municipality, addrly_lat, addrly_lon, addrly_similarity and addrly_status (ok, review, not_found, empty). It sends 100 rows per round trip and never spends a request on an empty cell.

Enterprise numbers, checked without a round trip

from addrly_api import is_valid_enterprise_number, normalize_enterprise_number

normalize_enterprise_number("BE 0123.456.749")  # "0123.456.749"
is_valid_enterprise_number("0123456748")  # False (mod-97)

Every company method accepts any format and validates locally first, so a typo raises InvalidEnterpriseNumber instead of costing a call.

Forward compatible on purpose

Models are plain dataclasses and every one keeps the payload it was built from in .raw. An unknown field never raises, so a field the API adds tomorrow is readable today:

company.raw["a_field_shipped_after_this_release"]

Only the stable shapes are typed. The deep dossier tree (annual-account snapshots, gazette events, participations) stays as plain dicts under typed entry points: those grow with the pipelines behind them, and a class per leaf would be wrong one release later.

Configuration

One client per process, shared between threads: it holds a pooled TLS connection, which is most of the latency of a first call.

Links

• API documentation: https://addrly.be/documentation
• Python guide: https://addrly.be/documentation/python
• llms.txt (paste into an AI coding assistant): https://addrly.be/llms.txt

MIT licensed. Belgium only, on purpose.