Python

Server SDKRuns on your backend (Calling Key)

The Python SDK automatically carries the call-time user and feature context on gateway requests. Tokens, cost, latency, and the request identifier are recorded by the gateway.

import os

from abto import init_abto

abto = init_abto(
    api_key=os.environ["ABTO_CALLING_KEY"],
    gateway_base_url="https://gateway.abto.app/v1",
    provider_keys={
        "openai": os.environ["OPENAI_API_KEY"],
        # Pass candidate keys together when the project routes across providers.
        # "anthropic": os.environ["ANTHROPIC_API_KEY"],
        # "gemini": os.environ["GEMINI_API_KEY"],
    },
)

openai = abto.openai()

with abto.with_context(
    device_id="device-abc",
    feature_id="review.summary",
):
    response = openai.chat.completions.with_raw_response.create(
        model="gpt-4.1-mini",
        messages=[{"role": "user", "content": "Summarize these 12 product reviews in three lines."}],
    )
    request_id = response.headers.get("x-abto-request-id")
    completion = response.parse()

completion: the usual OpenAI response body
raw response: where you read the x-abto-request-id the Gateway issued
gateway_base_url: required. Omit it and the SDK reads ABTO_GATEWAY_BASE_URL; with neither it raises ValueError

Gateway headers sent on every request

SDK input	Gateway header	Meaning
`api_key`	`Authorization: Bearer …`	ABTO Calling Key
`provider_keys["openai"]`	`x-abto-key-openai`	Your own OpenAI provider key
`feature_id`	`x-abto-feature-id`	Feature ID
`device_id`	`x-abto-device-id`	End-user device identifier (optional)

How provider_keys behaves:

Not a setting that registers keys with the Gateway. It is the input that attaches server-held credentials as per-request headers.
Passed with request scope only and never stored in call records.
If the routed provider has no key, the Gateway rejects that request.
Pass a callable instead of a string to re-evaluate it on every request. for key rotation and per-provider credential resolvers.

How caller-supplied headers are handled:

Do not override x-abto-key-* through extra_headers on a client built by abto.openai().
The Python SDK strips the caller’s Authorization, x-abto-key-*, x-abto-feature-id, and x-abto-device-id, then rebuilds them from the trusted init_abto settings and context.
Using the official OpenAI SDK directly instead? Put the same headers in each request’s extra_headers. see the direct OpenAI SDK example.

Context fields

The fields mean the same as in Node / Server JavaScript.

feature_id: feature ID (e.g. review.summary), sent as x-abto-feature-id
device_id: sent as x-abto-device-id. Take the browser-generated device_id from your backend request and pass it through so product behavior joins.

Rules for handling identifiers:

Validate client-supplied device_id and feature_id with your application’s existing request schema.
Never accept the Calling Key or provider keys from client requests. Read them only from server environment variables or a server-only credential resolver.

Joining responses to browser and mobile behavior:

Read x-abto-request-id with with_raw_response as shown above.
Include the same header in your backend response to join the Browser or Mobile SDK LLM trace.

OpenAI requests the Gateway supports

The data path is currently OpenAI Chat Completions.
Inline base64 images and PDFs in user messages are supported.
Unsupported fields such as streaming, tool calling, remote image URLs, and audio are rejected with 400 rather than silently ignored.
See Gateway OpenAI compatibility for the exact field list.

Retries happen at two layers

Because there are two layers, max_retries alone does not determine how many times a provider is called.

Layer	What it counts	Who decides
Client	Application → Gateway round trips	The official OpenAI SDK’s `max_retries`
Gateway	Gateway → provider invocations	Gateway built-in caps and the node retry policy

Client layer: `max_retries`

openai = abto.openai(max_retries=2)

Keeps the official OpenAI meaning: retries after the initial request. 0 is one round trip, 1 is two.
The Python SDK never overwrites it and offers no separate fallback retry count.
Leave it unset and the official OpenAI SDK default applies.
Every other official OpenAI option is forwarded unchanged.
api_key, base_url, and http_client are owned by ABTO for trusted routing. passing them raises ValueError rather than being silently ignored.

Gateway layer: retries inside a single round trip

Within one round trip the Gateway may re-attempt along the same path.

Network retries: only when non-delivery is certain (unreachable). Always on, regardless of node policy, up to 2.
Provider retries: 429 (rate limit), 500, 502, 503, 504, 529. Only when the node retry policy is enabled, up to 2.
The two budgets add independently, so one round trip can invoke the provider up to 5 times (1 initial + 2 network + 2 provider).
The x-abto-attempt response header reports which attempt produced the response.

Never retried:

429 with credit exhaustion (insufficient_quota). a deterministic failure
Timeouts, post-send disconnects, body size overruns, and cancellation. risk of double execution and double billing
Other deterministic statuses such as 501 and 505

Wait times:

Exponential backoff with full jitter. Network: 50ms base, 250ms cap. Provider: 500ms base, 8s cap.
A provider-supplied Retry-After is honored as given; beyond the 8s cap the Gateway does not wait and surfaces the error immediately.

OpenAI direct fallback on Gateway failures

This is the escape hatch back to the endpoint this application used before ABTO. Name that destination in fallback.base_url; there is no default. If you took an API key straight from OpenAI and used the official SDK, that address is https://api.openai.com/v1.

from abto import OpenAIDirectFallbackOptions, init_abto

abto = init_abto(
    api_key=os.environ["ABTO_CALLING_KEY"],
    gateway_base_url="https://gateway.abto.app/v1",
    provider_keys={"openai": os.environ["OPENAI_API_KEY"]},
    fallback=OpenAIDirectFallbackOptions(
        # The address this code called before ABTO was put in front of it.
        base_url="https://api.openai.com/v1",
        timeout_seconds=30,
        on_timeout=False,
    ),
)

The original Chat Completions body and model go straight to that address; the Gateway’s provider/model policy is not reproduced.

Sends the current request directly

Failures before the connection is established
Admission 503 raised before the provider call

Does not fall back the current request

Timeouts and disconnects with ambiguous delivery. the Gateway may already have run the provider
Provider, transport, and internal errors; deterministic 4xx and 429; anything after streaming has started
The direct circuit stays closed in these cases, and if the official OpenAI SDK retries it calls the Gateway again

Settings

base_url: required. The OpenAI-compatible endpoint used before ABTO. It must accept the OpenAI request path and Authorization: Bearer. Enabling fallback without it makes init_abto raise ValueError.
timeout_seconds: per-stage inactivity cap for Gateway connection pool wait, connect, write, and response header read. The body after headers and direct requests keep the abto.openai(timeout=...) value.
on_timeout=True: also resends the timed-out request. An explicit choice that accepts duplicate execution and billing risk.
fallback=False: turns the whole feature off.

Boundaries

Per SDK attempt the transport makes the Gateway judgment and the direct send exactly once each.
Direct calls bypass Gateway policy, ABTO telemetry, and request_id, and must use a model that endpoint supports.
Only headers OpenAI needs are forwarded; cookies, proxy credentials, and custom Gateway headers are stripped.
After switching to direct, OpenAI responses and errors are returned to the official OpenAI SDK unchanged, and that SDK decides whether to retry.
Native direct fallback for Anthropic and Gemini is out of scope today.