Connection guides

Detailed setup guides for SDKs, coding agents, IDEs, and OpenAI-compatible clients.

Choose the client you use:

Create an API key in the Model Gate control panel. The full key is shown only once.

Playground access

Open Playground in the account menu to use Request builder or Model Verification. On Business profiles the section remains visible while Model API is enabled, even when no usable key has been assigned yet. A setup page explains missing account verification, company activation, key permissions or API endpoint configuration and offers permitted management links and Support. Simply opening Playground makes no paid request and does not create a key.

Requests require a permitted active key. A project-bound key must belong to the current company and an active project. With developer or administrator execution access, a member may use a shared project credential in Playground without receiving its permanent secret. The browser session represents the signed-in member, not the key's custodian; viewing usage alone does not grant execution. When Model API is disabled for a domain, its Playground routes are also unavailable. Temporary-password and required two-factor setup still take precedence.

The panel uses its current domain profile's configured API endpoint. It does not switch to another project's API when a binding is missing. The key selector shows resettable usage against the configured spending limit; lifetime spend is not that counter. Current key/project limits and company balance are checked before issuing a short-lived session, and the API remains authoritative for live admission and billing. A project freeze or permission change after page load can make the previously selected key unavailable; refresh the page after an authorized correction.

Multimodal request formats

The Builder follows the same capability contracts used by administrator model probes. OpenAI-compatible image generation uses /v1/responses with the hosted image_generation tool; the generated image_generation_call.result is decoded and shown as an image artifact. OpenAI Responses PDF input uses an input_file content item with a complete data:application/pdf;base64,... value in file_data. Anthropic Messages PDF input uses a native document block with source.type=base64, media_type=application/pdf and the bare base64 bytes. The browser reads selected files locally and sends them directly to the configured Model API endpoint; the panel does not upload them first. The Builder does not advertise Chat Completions as a PDF format. Raw JSON remains available for provider-specific contracts outside these tested shapes.

Browser Model Verification

/chat/verify compares an endpoint with a ready behavioral reference of the selected canonical model. The current method is model-verification-behavioral-reference-v2.0. It is not a provider attestation or a claim about training records.

Ready reference, no research stage

Administrators separately enable a paid reference collector using an explicitly trusted route. It reads ready months from the existing question bank, with ten judged observations per question by default and up to twenty distinct questions per collection request. Provider failures and ungraded/missing answers are not observations of ignorance. Administrative answers and revisions are retained for inspection; user runs never contribute to this trusted reference automatically.

A reference must cover sixteen consecutive completed months on both sides of a stable observed RECALL anchor. Two eligible facts are selected per month, preferring one usually-correct and one usually-not-correct fact when both exist; otherwise two of the available stable category. A 32-question format-control call must be accepted by the judge before the reference becomes ready. Enabled automatic collection queues it under the shared API-call authorization; legacy manual profiles retain their explicit control action. Official cutoff metadata remains separate. An unpublished official cutoff never triggers research during a user run. No ready reference or incomplete coverage means the model cannot be tested yet.

One tested request and one foreground judge

Select a ready reference and an eligible Model Gate key for evaluation billing. The reference pins the protocol, canonical model and exact request contract. New automatic references use minimal-v1: Chat sends only model + one user message, Responses only model + input, and Anthropic Messages additionally sends required max_tokens=4096. Historical references keep their saved legacy options. The browser makes one request with 32 questions to your selected HTTPS endpoint. The external key remains only in browser memory; it is never sent to the Model Gate application server or stored in local/session storage. CORS must permit that browser request. The tested model receives questions and closed-book instructions, never the answer key, reference answers, positive/negative labels or expected mistakes.

There is no cutoff lookup, hosted search, bank population or question generation in this lifecycle. Local snapshot selection is not a paid generation call. The external provider may charge for its request; the foreground judge uses the selected eligible Model Gate credential. Keep the tab open. The live stage indicator shows elapsed time and actual work; while a verification is active, the browser warns before close, reload or navigation. This does not continue the request in the background. Retrying an incomplete evaluation uses the answer still held in the browser, not a second tested-model request. An uncertain call is not automatically replayed.

The browser streams the foreground judge Responses request directly through api.model-gate.com, without hosted search. The multi-minute judge stream is not held open through panel.model-gate.com. Automatic format checks and the old Instruction Following score remain excluded; semantic assessment belongs exclusively to the judge.

Judge-only correspondence

The judge receives the current answers, correct answers and all grouped reference observations for the selected 32 facts, including contradictory observations and their multiplicities. It evaluates correctness, declared RECALL/INFERENCE/GUESS/UNKNOWN behavior and departures in either direction. Unexpected correct answers to usually-unsolved reference questions can reduce correspondence; actual correct answers are not relabeled wrong. The judge alone assigns the correspondence score, identity-assessment category, confidence, reasons and limitations. The server validates report structure and counts saved observations but applies no semantic threshold, penalty, weighted score or replacement conclusion.

A high score is not a probability of exact model identity. The report cannot establish a model's price or training date. The stored snapshot and benchmark identify exactly what was compared. Old v1.7 reports retain their original method and are not merged with v2 evidence. No automatic community identity badge is computed from v2 scores.

Reports and privacy

The live trace shows the request, browser-held response, evaluation payload and judge output without credentials. A private run does not retain the full raw provider response or per-question verbatim visible answer for browser recovery; normalized grades, explanations and immutable reference identifiers remain. Publication is a separate explicit opt-in. Published v2 reports display the saved judge conclusion without running another evaluation or server-side rescoring. Administrative reference answers have their own retained/audited lifecycle, unlike ordinary browser responses.

Request compatibility and provider errors (10.28.1)

The Chat Completions verification request places the complete closed-book instructions and all questions, in order, into one user message. This avoids requiring an OpenAI-compatible gateway to translate a system role when it routes to an Anthropic-style backend. The selected protocol and exact endpoint remain authoritative: there is no model-name-based protocol switch, extra tested-model request or general-purpose proxy conversion. Native Anthropic Messages still uses top-level system with a user message; Responses retains its canonical user input. The trace shows the actual outbound JSON. The established transport compatibility is retained. Behavioral-reference v2 supplies a pinned 32-question request without changing the one-tested-model-request rule.

Non-2xx responses are HTTP failures, not unrecognized successful envelopes. A root provider error is reported as Provider error envelope, even when returned with HTTP 200. Such errors never supply model answers, never reach judge evaluation, and cannot be repaired by Retry evaluation of the same response. Protocol compatibility is NOT_ASSESSED for an error response. Any executed external request may still be billable; local reference selection makes no generator call. Raw provider errors remain browser-local diagnostics rather than public report content. A generic Model Gate upstream error deliberately does not reveal the administrator-only raw upstream message.

Automatic reference administration

Administrators configure automatic reference collection once for eligible existing and future catalog models. The configured Model Gate API connection and administrator key are used, with shared daily/lifetime API-call limits and no separate automatic USD reservation budget. Models need an official cutoff or a separately sourced unofficial estimate to anchor collection; the worker does not research missing dates or generate questions. The initial default remains ten usable observations per question and at most twenty different questions per collection request. Legacy manual references remain separate.

The reference is evidence of configured behavior, not proof of which provider executed a request. Ineligible or ambiguous routes remain unavailable until the administrator resolves routing. Only the judge assesses correspondence.

Unsupported API paths now return 404 endpoint_not_supported with the requested path; a wrong method on a supported path returns 405 method_not_allowed and Allow. Native /v1/balance is unchanged; /v1/user/balance is not implemented. Share the returned request ID, never your API key, when reporting these errors.