neutral-llm-gateway

A small, honest gateway for LLM calls: typed contracts, thin provider adapters, and accounting of retries, fallbacks, tokens and cost that refuses to lie to you.

It knows about providers, never about products. It holds no credentials, reads no environment variables, ships no prompts and stores no business schemas. Everything product-shaped — ledgers, tenants, alerting, history, prompts — stays in your application, wired in through optional ports.

your application → your facade → llm_gateway → provider SDK

Never the reverse.

Why this exists

It was extracted from several applications that had each grown their own version of the same "call an LLM" function — the largest close to two thousand lines, mixing provider calls, retry policy, cost maths, a usage ledger and business alerting in one place. Writing that from scratch a fourth time is how subtle accounting bugs get copied around.

The bugs it is built to prevent are all the same shape: a number that looks like a fact but isn't.

• A provider returns no usage, the code records 0 tokens, and the call bills as free.
• A model is missing from the price table, so its cost is USD 0.00 — indistinguishable from a genuinely free call.
• A retry fails after the model produced tokens, and only the successful attempt gets counted.
• A fallback quietly answers with a different model, and the metrics attribute it to the one you asked for.

Here, unreported usage is None, unknown cost is UNAVAILABLE, every billable attempt is counted, and a fallback is never silent.

Is this for you?

Probably yes if you want a thin, auditable layer you can read in an afternoon, you want to own your credentials, you care about per-call cost being reconcilable against an invoice, and you prefer typed results to dictionaries.

Probably not if you want routing across dozens of providers, a proxy server, streaming, or an agent framework. LiteLLM and friends cover far more surface than this does. This covers deliberately less, and is explicit about what it does not know.

Install

Provider SDKs are optional extras. Install only what you call.

# uv
uv add "neutral-llm-gateway[gemini]"

# pip
pip install "neutral-llm-gateway[gemini]"

Available extras: openai, gemini, groq, assemblyai, openrouter, replicate, wavespeed, all. Combine them as [openai,assemblyai]. openrouter installs the openai SDK, and assemblyai and wavespeed install the small HTTP transport used by their REST adapters.

Importing the package with no extra installed works by design; asking for a provider you have not installed raises a typed error naming the exact extra.

To try an unreleased commit, the PEP 508 git form works everywhere and pins a tag or revision:

pip install "neutral-llm-gateway[gemini] @ git+https://github.com/jmgb/llm-gateway-python.git@v0.5.0"

Use

from pydantic import BaseModel

from llm_gateway import (
    FallbackPolicy,
    LLMGateway,
    LLMRequest,
    Message,
    ResponseFormat,
    RetryPolicy,
)
from llm_gateway.factories import build_registry, create_gemini_client


class Answer(BaseModel):
    verdict: str


# You build the client, so you keep the key. Prices come from the built-in
# versioned catalogue unless you pass your own.
gateway = LLMGateway(
    registry=build_registry(gemini_client=create_gemini_client(api_key=my_key)),
)

result = await gateway.generate(
    LLMRequest(
        model="gemini-3.5-flash-lite",
        system_prompt="Answer strictly from the supplied evidence.",
        messages=(Message("user", question),),
        response_format=ResponseFormat.JSON_SCHEMA,
        response_schema=Answer,
        temperature=0,
        retry_policy=RetryPolicy.transient(max_attempts=2),
        fallback_policy=FallbackPolicy.disabled(),
        request_id=request_id,
        source="my-feature",
    )
)

result.output  # Answer instance — no metadata mixed in
result.usage.input_tokens  # None means "not reported", not zero
result.cost.amount_usd  # None when unavailable, never a fake 0
result.cost.measurement  # ACTUAL | ESTIMATED | UNAVAILABLE
result.execution.model_used  # provider-reported model that actually answered
result.execution.fallback_used  # whether the gateway used its fallback plan
result.execution.attempts  # every attempt, including the failed ones

If the model answers with something that is not valid JSON, or with JSON that violates Answer, that attempt is recorded as failed and billed — the tokens were spent — and the next model in fallback_policy is tried. When no model produces a usable answer the call raises AllAttemptsFailed, carrying every attempt, with the parsing or schema error as its __cause__. Schema validation errors name each Pydantic loc and type, while dynamic response keys and values stay out of the message.

Tool calling

The gateway declares your functions, hands you the calls the model made, and puts your results back on the wire. It never runs a function: authorisation, side effects and business schemas are yours, and so is the loop.

from dataclasses import replace

from llm_gateway import FunctionTool, LLMRequest, Message, ToolResult

request = LLMRequest(
    model="gpt-5.6",
    messages=(Message("user", "What is the weather in Madrid?"),),
    tools=(
        FunctionTool(
            name="get_weather",
            description="Current weather for a city",
            parameters={
                "type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"],
            },
        ),
    ),
)
result = await gateway.generate(request)

if result.tool_calls:
    call = result.tool_calls[0]
    output = await my_app.run(call.name, call.arguments)  # yours, not the package's
    result = await gateway.generate(replace(request, tool_results=(ToolResult(call, output),)))

result.output  # the answer, once the model stopped asking for functions

tool_choice takes ToolChoice.AUTO, NONE, REQUIRED or RequiredTool("get_weather") to force one. A ToolResult holds the ToolCall it answers rather than a loose id, because both halves go back on the wire and a pair that drifted apart answers the wrong question.

Tool calls never arrive as output: when the model calls a function output is None and tool_calls is non-empty, and a requested JSON format is not parsed out of a reply that contains no answer. Arguments the caller could not dispatch — unparseable, not a JSON object, or naming a function the request never declared — make that attempt a billed failure, so the fallback still gets its turn and the money is still counted. The package checks what makes a call dispatchable, not the whole JSON Schema: that needs a validator this package does not depend on, and the code that owns the function validates anyway. Definitions, arguments and results never reach a usage, event or alert sink.

Transcription

Speech-to-text is a separate operation with duration usage and audio pricing:

from llm_gateway import AudioInput, LLMGateway, TranscriptionRequest
from llm_gateway.factories import build_registry, create_openai_client

gateway = LLMGateway(
    registry=build_registry(openai_client=create_openai_client(api_key=my_key)),
)
transcript = await gateway.transcribe(
    TranscriptionRequest(
        model="gpt-transcribe",
        audio=AudioInput(
            data=audio_bytes,
            filename="voice.webm",
            mime_type="audio/webm",
            duration_seconds=12.5,
        ),
        language="es",
        source="voice-note",
    )
)

transcript.text
transcript.usage.duration_seconds
transcript.cost.amount_usd  # audio minutes, never token pricing

language is optional and defaults to provider detection. Use assemblyai-universal-3-pro or assemblyai-universal-2 with a public AudioInput.url, and whisper-large-v3-turbo/whisper-large-v3 with Groq. AssemblyAI Universal-3 Pro supports prompt and speaker labels; OpenAI and Groq reject speaker labels rather than ignoring them. Fallbacks are explicit through FallbackPolicy.models_in_order(...).

Image generation

Images are a separate operation too, for the same reason: the output is bytes or a URL, and providers bill it either per image or with image-output tokens whose rate can differ from text output.

from llm_gateway import ImageInput, ImageRequest, LLMGateway
from llm_gateway.factories import build_registry, create_gemini_client

gateway = LLMGateway(
    registry=build_registry(gemini_client=create_gemini_client(api_key=my_key)),
)
result = await gateway.generate_image(
    ImageRequest(
        model="gemini-3.1-flash-image",
        prompt="a cat wearing a hat, studio lighting",
        image=ImageInput(data=original_bytes, mime_type="image/jpeg"),  # optional: an edit
        source="whatsapp-image",
    )
)

result.images[0].data  # bytes from Gemini; `.url` from Replicate and WaveSpeed
result.usage.images
result.cost.amount_usd  # per image, or per token where the model bills that way

Which form comes back is the provider's, not a choice: Gemini returns inline bytes, Replicate and WaveSpeed return a URL. Downloading it, re-hosting it, watermarking it and enforcing a per-user quota stay in the application, exactly as decoding audio does.

Editing needs a source image in the form the provider accepts — bytes for Gemini, a URL for Replicate — and an adapter that cannot use the form it was given raises rather than dropping it. WaveSpeed's text-to-image models refuse an edit outright. Sending an image model through generate() raises too: its reply carries no text, and returning it as an empty success is the failure this separation exists to prevent.

Video generation

Video is a third seam, billed by the second:

from llm_gateway import ImageInput, LLMGateway, VideoRequest
from llm_gateway.factories import build_registry, create_wavespeed_client

gateway = LLMGateway(
    registry=build_registry(wavespeed_client=create_wavespeed_client(api_key=my_key)),
)
result = await gateway.generate_video(
    VideoRequest(
        model="wavespeed-ai/minimax-h3/image-to-video",
        prompt="the lioness sprints and leaps at the gazelle",
        image=ImageInput(data=first_frame_bytes, mime_type="image/png"),
        resolution="480p",
        duration_seconds=5,
        source="wildlife-clip",
    )
)

result.videos[0].url
result.usage.seconds, result.usage.resolution
result.cost.amount_usd  # per second, at the rate for that resolution

The first frame may be bytes or a URL — WaveSpeed accepts an inline data URI, which is what lets an image from one provider be animated by another without hosting it anywhere first.

A clip takes minutes, so VideoRequest defaults to a fifteen-minute total budget and the adapter owns the provider's polling loop. Resolution is part of usage rather than a detail of the request because the rate depends on it: a resolution the price table does not know costs UNAVAILABLE, never the cheaper rate.

Omit resolution and you get the cheapest tier the model offers, never the provider's own default — which is dearer on every catalogued video model. It is the biggest single lever on a video bill (480p vs 768p is 2× on MiniMax H3), so silence buys the floor and anything above it is asked for by name. See the table in docs/pricing.md.

Video jobs, for providers that answer minutes later

Replicate does not hand back a clip; it hands back a prediction id and finishes later. Waiting inside one await would mean a four-minute request timeout and no way to use the webhook, so the same VideoRequest takes a second route:

from llm_gateway import LLMGateway, VideoJob, VideoJobStatus, VideoRequest
from llm_gateway.factories import build_registry, create_replicate_client

gateway = LLMGateway(registry=build_registry(replicate_client=create_replicate_client()))

job = await gateway.submit_video(
    VideoRequest(
        model="wan-video/wan-2.2-5b-fast",
        prompt="the lioness sprints and leaps at the gazelle",
        image=ImageInput(url="https://cdn.example/first-frame.png"),
        webhook_url="https://app.example/hooks/video",  # optional
    )
)

job.id, job.model, job.provider, job.status  # required provider state
job.request_id, job.source  # optional accounting correlation: store them too

VideoJob is plain data on purpose — the process that polls is usually not the one that submitted. Save its fields, rebuild it later, and read it back from anywhere:

result = await gateway.poll_video(
    VideoJob(
        id=row.id,
        model=row.model,
        provider=row.provider,
        request_id=row.request_id,  # so the eventual cost stays attributable
    ),
    timeout_seconds=30.0,  # bounds the status call, not the job
)

if result.job.status is VideoJobStatus.SUCCEEDED:
    result.videos[0].url
    result.usage.seconds  # measured by the provider, where it reports one
    result.cost.amount_usd
elif result.job.status is VideoJobStatus.FAILED:
    result.error  # why the provider gave up, when it says

request_id and source are copied from the VideoRequest into the job, so store them alongside the rest: the clip is billed minutes later from another process, and they are what ties that amount back to the call that caused it.

A job the provider gave up on is a status, not an exception: the submission worked, and an application storing the outcome wants the reason rather than a traceback. poll_video() raises only when the reading failed — including when the status call itself exceeds timeout_seconds, which stops a stalled provider from blocking the worker that polled it.

Nothing is billed until the job is terminal. A submission records no usage — the clip does not exist yet — and a job polled ten times is recorded once.

Three Replicate video models are catalogued, and the adapter translates each one from its published input schema rather than a shared guess:

That last column is the cheapest tier of each, and it is deliberately not what Replicate would have picked — its own defaults are 720p, pro (1080p) and 720p respectively.

A resolution a model does not offer raises rather than falling back to its default: Replicate ignores keys it does not recognise, generates the default, and bills for it, so a silent drop is a clip nobody asked for on the invoice. None of the three has a per-second rate this package could verify, so their cost is UNAVAILABLE — pass a VideoPriceCatalog with your measured rates.

The guarantees

These are enforced by tests, not by convention:

Extending

Ports are optional and default to no-op: UsageSink, EventSink, AlertSink, AudioUsageSink, PriceCatalog and AudioPriceCatalog. Implement what you need; the package will not reach into your application to find them.

Adding to the public API follows the two-consumer rule: nothing is promoted into the core until two distinct applications need it. Until then it belongs in that application's local adapter.

Providers

Capabilities are declared per provider and never faked as identical — query adapter.capabilities before relying on one.

A provider that declares structured_outputs=False has no API field that binds the answer to a shape, so its adapter states the requested schema in the system prompt instead of dropping it. Otherwise the model answers valid JSON under keys of its own choosing, validation rejects it, the attempt is billed anyway and the fallback serves every structured call — a result that looks correct and shows up only on the invoice.

The same adapters add one sentence asking for JSON when JSON_OBJECT is requested, because Groq rejects that mode with HTTP 400 unless the word appears in the messages. Setting response_format is what creates the obligation, so the adapter meets it rather than the caller's prompt — and a prompt that already says "json" is left as it is.

OpenAI and Groq declare function_calling=True, each in its own dialect: the Responses API takes a flat tool and answers with function_call items carrying a call_id, while Chat Completions nests the function and answers inside choices[0].message.tool_calls. Groq sends tools or a JSON response_format and never both, which costs nothing — it enforces no schema either way, the system prompt still states the shape and the gateway still validates. Gemini and OpenRouter declare False and reject a request carrying tools rather than answering prose where a call was expected; their request, response and continuation shapes get the capability once deterministic tests cover them.

inline_files remains unsupported. OpenAI declares remote_files=True: LLMRequest.attachments accepts already-uploaded file IDs and the Responses adapter appends them to the last user message as input_file parts. Providers without that capability reject the request rather than silently dropping the files. OpenAI, Groq and AssemblyAI expose audio_transcription=True through the separate TranscriptionRequest API, and Gemini, Replicate and WaveSpeed expose image_generation=True through ImageRequest. WaveSpeed and Replicate both declare video_generation=True and video_from_image=True, reached through VideoRequest — WaveSpeed by awaiting generate_video(), Replicate through submit_video() and poll_video(). Only Replicate declares video_webhooks=True, so an application can ask whether it must poll before committing to a design.

Request options are adapted per model before each API attempt. A model that rejects temperature — the OpenAI 5.6 family — never receives it, including when it is reached through a fallback that inherited it from another model.

Reasoning effort is checked the same way. OpenAI 5.6 models support none, low, medium, high, xhigh, and max; Gemini 3 Flash supports minimal, low, medium, and high; Gemini 3 Pro and Groq GPT-OSS support low, medium, and high. If a fallback cannot honour the requested effort, the gateway uses medium when available and otherwise omits the reasoning option.

LLMRequest.verbosity (low, medium, high) asks for a shorter or longer answer, which is not what max_output_tokens does: that one truncates an answer already being written and pays for every token up to the cut. OpenAI declares verbosity=True and sends it for the gpt-5 families; an adapter without the capability has no field for it, so a fallback that inherits the option is neither charged for it nor broken by it.

LLMRequest.routing states which upstream should serve the call. Only an aggregator has that choice, so OpenRouter is the one adapter declaring upstream_routing=True:

from llm_gateway import LLMRequest, RoutingPreference

LLMRequest(
    model="google/gemini-3.6-flash",
    routing=RoutingPreference(order=("Groq", "SambaNova"), optimise_for="throughput"),
)

Stating nothing is the default, and it is not the same as stating an empty preference: a blank instruction would override the aggregator's own routing, so only the halves a caller filled in are sent.

[openrouter] installs the openai SDK, because OpenRouter speaks the OpenAI wire format and ships none of its own. That is a fact about the transport: the adapter, the declared capabilities and the prices are OpenRouter's.

Routing to OpenRouter

Models reach it by their namespace, so nothing needs configuring:

from llm_gateway.factories import (
    build_registry,
    create_openai_client,
    create_openrouter_client,
)

registry = build_registry(
    openai_client=create_openai_client(api_key=...),
    openrouter_client=create_openrouter_client(api_key=...),
)

registry.resolve("gpt-5.6-luna")  # openai
registry.resolve("deepseek/deepseek-v4-pro")  # openrouter, from the catalogue
registry.resolve("somevendor/brand-new")  # raises: not in the catalogue
registry.resolve("openai/gpt-oss-120b")  # groq — the prefix is not OpenAI

gemini-3.1-pro-preview and google/gemini-3.1-pro-preview are the same model on two routes, catalogued separately because they are billed separately.

For any other OpenAI-compatible endpoint — Azure, vLLM, your own gateway — pass base_url to create_openai_client and widen the routing with build_registry(extra_openai_prefixes=...).

Model catalogue and prices

The package ships a versioned table of models — provider, and price in USD per million tokens — used by default, so a call is priced without you wiring anything up. Audio models such as gpt-transcribe, Whisper and AssemblyAI are catalogued with their duration unit and priced through a separate AudioPriceCatalog; image models carry their per-image rate, or their token rate where the provider bills images as tokens, and are priced through an ImagePriceCatalog; video models carry a per-second rate that depends on the resolution and are priced through a VideoPriceCatalog. Override prices for negotiated rates, or implement either catalog protocol yourself. See docs/pricing.md.

Not in this version

Inline files, streaming and Gemini File Search remain absent, and so do provider-hosted tools — web search, code execution, file search — and any package-owned loop that would execute a function for you. Function tools, remote file IDs for OpenAI, audio transcription, image generation and video generation — both the polled shape and the submit/poll job shape — are supported with their own capability and cost contracts.

Documentation

• docs/architecture.md — how the split was decided
• docs/pricing.md — cost model and updating prices
• docs/migration.md — adopting it behind an existing function
• CONTRIBUTING.md — what belongs here, and the non-negotiables

Releasing without GitHub Actions

The local release runner keeps versioning and publication independent from GitHub Actions minutes. Preview a release first:

uv run --offline python scripts/release.py --version 0.6.0 --dry-run

Prepare the release locally, including tests, the version in pyproject.toml and uv.lock, the changelog, a release commit, and an annotated tag:

uv run --offline python scripts/release.py --version 0.6.0

Add --push to push main and the tag. Add --publish as well to publish the matching wheel and sdist with uv publish and create the GitHub Release; the latter requires GitHub CLI authentication and UV_PUBLISH_TOKEN. --publish implies a real external release and therefore requires --push.

Every built artifact is audited before it is uploaded, and the release is refused if the archive contains an unexpected dotfile or a credential-shaped name. What reaches a package index cannot be recalled — the file is mirrored within minutes — so the check runs between the build and the upload, which is the last moment it is still worth anything.

The local runner is the normal publisher when Actions minutes are unavailable. The GitHub workflow is manual only; use one publisher per version to avoid uploading the same PyPI files twice.

Development

uv sync
uv run pytest        # no network, no cost, no extras required
uv run ruff check .
uv run mypy
uv build

Python 3.11+, Pydantic v2.

Status

0.x: in production use, but the API may still change between minor versions. Pin an exact version. Every release documents its changes, and cost-affecting changes are called out explicitly.

License

MIT.