Arbitr Python

Official Python client and arbitr CLI for the Arbitr External API.

Install

pip install arbitr-sdk

The PyPI package is arbitr-sdk. The import and CLI stay arbitr. Python 3.11 or newer.

Mint a key at https://app.arbitr.ai/settings/api-keys and store it as ARBITR_API_KEY in the environment or a .env file. Submit and download need both verify:submit and verify:read. A test key (abr_test_...) checks auth and request shape without running the pipeline or spending credits — start there, then switch to a live key (abr_live_...). See Test mode.

Library

from arbitr import ArbitrClient

client = ArbitrClient.from_env()  # ARBITR_API_KEY, optional ARBITR_BASE_URL
project = client.projects.submit(
    files=["report.docx"],
    name="Q3 report",
    target_language_codes=["ko-kr", "fr-fr"],
    idempotency_key="run-2026-08-report",
)
final = client.projects.wait(project.id)
client.projects.download_zip(final.id, "out/deliverables.zip")

Async:

from arbitr import AsyncArbitrClient

async with AsyncArbitrClient.from_env() as client:
    me = await client.me()

Default host is production: https://api.arbitr.ai, deep-linking to the UI at https://app.arbitr.ai. The former *-arbitr.straker.ai hosts redirect; the client does not follow redirects, so drop any base_url override still naming them.

Language codes are lowercase BCP-47 tags (ko-kr, fr-fr). Bare codes (ko) are rejected by the API — client.languages.resolve(["ko"]) expands them, or arbitr submit --resolve-locales does it for you.

The client wraps the published OpenAPI surface only. Former aliases (agent-selection, /deliverables/zip, /resume) have been removed from the API and answer 410 Gone; this package already uses the canonical replacements (wait() / the Arbitr UI, ?format=zip, /resumptions).

CLI

export ARBITR_API_KEY=abr_live_...
arbitr me
arbitr submit report.docx --locales ko-kr,fr-fr --wait --out out/

arbitr --help and arbitr COMMAND --help are the command reference.

Exit codes:

Coding agents

Give your agent this prompt:

Set up arbitr for me. Fetch https://arbitr.apidocumentation.com/getting-started/agent-setup/index.md and follow it.

It installs the CLI, pauses while you create and store a key, verifies access with arbitr me, and installs the persistent skill. It does not submit a project.

Or install the agent skill directly (source: skills/arbitr/SKILL.md):

npx -y skills add strakergroup/arbitr-python --skill arbitr --yes --global

To call the API as hosted tools instead of through the CLI, see MCP.

Errors

Everything this package raises derives from ArbitrBaseError, so one except is enough to be safe. Below it there are two branches:

httpx exceptions never escape the client — connection and timeout failures arrive as ConnectionFailedError / RequestTimeoutError with the original exception on __cause__.

from arbitr import ArbitrBaseError, PaymentRequiredError

try:
    client.projects.resume(project_id)
except PaymentRequiredError as exc:
    print(f"short by {exc.shortfall} credits")
except ArbitrBaseError as exc:
    print(f"call failed: {exc}")

Retries and rate limits

Retries are opt-in and off by default on the library client:

client = ArbitrClient.from_env(max_retries=3)

The CLI retries GET up to 3 times (--max-retries / ARBITR_MAX_RETRIES).

Only GET is replayed — Retry-After is honoured on 429 and 5xx back off exponentially (capped at 60s). POST /v1/projects is never retried automatically because its multipart body streams file handles; pass idempotency_key= and retry it yourself. After any call, client.rate_limit holds the latest X-RateLimit-* values.

Contributing

Development setup, checks, and how to refresh the pinned OpenAPI spec are in CONTRIBUTING.md.