GateViaAI CONTROL PLANE
Create workspace
Private preview · staging-qualified

The control plane
for production AI.

GateVia gives engineering, finance and operations one system for multi-model inference: an OpenAI-compatible endpoint, a price-aware model catalog, workspace-scoped credentials, and a prepaid ledger behind every request.

Create a workspace Read the platform docs
OpenAI-compatible API SSE streaming Prepaid settlement Tenant isolation
MODELS30 provider models
PROVIDERS17 catalog authors
INTERFACEOpenAI chat-completions
STREAMINGServer-sent events
BILLINGPrepaid ledger
WHY GATEVIA

Model access is easy.
Operating it is not.

A production AI stack needs a catalog that matches the bill, credentials that do not leak, spend that can be audited, and a boundary between teams. GateVia makes those one system instead of four integrations.

01

One API contract

Keep the OpenAI client you already use. Put model and provider choice behind a single chat-completions endpoint instead of rebuilding integrations for every upstream.

02

One source of truth

Catalog rows, endpoint prices and settlement all read from the same published snapshot. The console does not invent a second price table.

03

One tenant boundary

Organizations own billing. Workspaces scope keys, usage and policy. An inference key from one workspace cannot act as a console session.

04

One audit trail

Usage rows, reservation and settlement records, request identifiers and operator events give finance and engineering the same evidence.

ARCHITECTURE

A request path you
can explain.

Each plane has one job. The console is not the financial authority. The gateway is not the catalog. Provider credentials never reach the browser.

CLIENTYour applicationOpenAI SDK or plain HTTP
EDGEGateVia APIBearer key · scope · rate limit
EXECUTIONModel gatewayRoute · provider grant · SSE proxy
UPSTREAMProvider endpointOpenRouter or BYOK
CATALOGPublished model + price snapshot
FINANCEReservation + immutable ledger
VAULTEncrypted provider credentials
CONSOLEKeys, usage, invoices, audit
Design invariants
  • The price book used for settlement is the catalog the customer sees.
  • Provider secrets stay server-side; the browser only handles workspace keys.
  • Missing analytics or health data is reported as missing, never zero-filled.
  • Routing intelligence can recommend; it cannot silently activate.
CAPABILITIES

Built as one system,
not a proxy.

The useful parts of a model gateway are not only the proxy hop. They are the controls around identity, price, credentials, money and evidence.

01API

OpenAI-compatible inference

Send chat-completions requests to /v1 or /api/v1. Stream deltas over SSE. Provider errors keep their original status and body.

  • Chat-completions wire format
  • [DONE] stream terminator
  • Same path used by the playground
02CATALOG

Price-aware model catalog

Browse context limits, supported parameters, provider endpoints and per-million prices. The router and the invoice read the same rows.

  • Input, output and cache pricing
  • Endpoint-level metadata
  • Published snapshot provenance
03IDENTITY

Workspace-scoped keys

Mint one-time inference keys that are stored only as SHA-256 hashes, scoped to a workspace and revocable without touching a session.

  • Optional key limits and allowed models
  • Console session and inference bearer are separate
  • Keys shown once, never returned again
04FINANCE

Prepaid finance with a ledger

Fund a workspace with invoice-backed prepaid settlement. Usage is metered after routing and recorded through the finance authority.

  • Credits, budgets and usage rows
  • Idempotent ledger entries
  • Refund and reconciliation workflows
05TENANCY

Organizations and roles

Organizations are billing scopes; workspaces are request scopes. Invite members and assign owner, admin, billing or viewer roles.

  • Multiple workspaces per account
  • Workspace-level key administration
  • Narrow database functions, no superuser DSN
06OPERATIONS

Operable from day one

Health, readiness, Prometheus metrics, graceful shutdown and signed release artifacts are part of the product, not an afterthought.

  • Readiness fails closed when a dependency is down
  • Backup, restore and failure-injection drills
  • Versioned artifact rollback rehearsal
MODEL CATALOG

The catalog is the contract.

Published model rows, context limits and endpoint prices come from the same snapshot used by routing and settlement. No separate marketing price list.

Open the full catalog
30 provider models 17 catalog authors Prices per 1M tokens Snapshot-backed
Popular

Anthropic: Claude Fable 5

anthropic
anthropic/claude-fable-5
Context1,000,000
Input / 1M$10.00
Output / 1M$50.00
Supports max_tokens, stop, temperature, top_p
DetailsOpen
Recommended

Anthropic: Claude Fable 5.1

anthropic
anthropic/claude-fable-5.1
Context1,000,000
Input / 1M$10.00
Output / 1M$50.00
Supports max_tokens, stop, temperature, top_p
DetailsOpen
Published

ByteDance Seed: Seed 1.6

bytedance-seed
bytedance-seed/seed-1.6
Context262,144
Input / 1M$0.25
Output / 1M$2.00
Supports max_tokens, stop, temperature, top_p
DetailsOpen
Published

ByteDance Seed: Seed 1.6 Flash

bytedance-seed
bytedance-seed/seed-1.6-flash
Context262,144
Input / 1M$0.07
Output / 1M$0.30
Supports max_tokens, stop, temperature, top_p
DetailsOpen
Published

Cohere: Command A

cohere
cohere/command-a
Context256,000
Input / 1M$2.50
Output / 1M$10.00
Supports max_tokens, stop, temperature, top_p
DetailsOpen
Published

DeepSeek: DeepSeek V4 Flash 0423

deepseek
deepseek/deepseek-v4-flash
Context1,048,576
Input / 1M$0.09
Output / 1M$0.18
Supports max_tokens, stop, temperature, top_p
DetailsOpen

Catalog authors include Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, Qwen, xAI, Cohere, Moonshot, MiniMax, NVIDIA and others. OpenRouter is the current pilot provider; BYOK storage is implemented, while other live provider onboarding remains a separate decision.

SECURITY & GOVERNANCE

Security is a boundary,
not a badge.

We do not claim certifications we do not hold. We do show the controls that are implemented, the evidence behind them, and the boundaries that remain open.

CREDENTIALS

Keys that cannot become sessions

Inference keys are workspace-scoped, shown once and stored as SHA-256 hashes. A console cookie cannot authenticate an inference request; an inference key cannot mint console actions.

PROVIDER SECRETS

Server-side by construction

Provider credentials are encrypted at rest and obtained through an opaque per-attempt vault grant. The raw provider secret is never returned to the browser.

MONEY

The product never moves a balance

The finance authority owns balances and ledger entries. GateVia requests settlement through narrow APIs, uses idempotent keys and preserves obligation history.

EVIDENCE

Missing data stays missing

Audit events, usage rows and request identifiers provide an evidence trail. When analytics or a dependency is unavailable, the console reports the gap instead of inventing a zero.

Current boundaries and non-claims
  • This is an isolated private preview. Production ingress, public DNS/TLS and production activation are not enabled.
  • Card/PSP processing is not activated. Funding is invoice-backed prepaid settlement through the ledger; there is no mock card flow.
  • Auto and Pareto routing remain shadow-only and default-off. Live activation requires a separate owner and operator decision.
  • SSO/SCIM, Responses/Anthropic API skins, image/video generation and real email delivery are not implemented in this slice.
  • No production SLA or certification is claimed. Qualification evidence comes from isolated staging and production-shaped rehearsal environments.
Read the developer and boundary documentation
QUALIFICATION

Qualified before
it is announced.

The platform is exercised as a system, not a collection of screens. These are current staging-qualification results, not a production guarantee.

37 / 0Product E2EFull customer journey
133 / 0Finance acceptanceLedger and settlement matrix
12 / 0Security negativesAbuse and boundary checks
17 / 17Failure injectionDegraded-path behavior
6.1sRestore RTOMatching source/restore counts
28 / 30Real provider matrixTwo provider-side failures open

Current open items include two provider-side model failures in the real OpenRouter matrix. Production external inputs — ingress, paging, off-host backup custody, KMS and production credentials — remain outstanding. These are reported, not hidden.

FIRST REQUESTSSE
curl "$GATEVIA_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $GATEVIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic-claude-fable-5",
    "messages": [{
      "role": "user",
      "content": "Summarize this incident."
    }],
    "stream": true
  }'

HTTP/1.1 200 OK
content-type: text/event-stream

data: {"choices":[{"delta":{"content":"..."}}]}
data: [DONE]
OpenAI-compatibleStreamingMeteredRequest IDs
DEVELOPER PATH

From workspace
to first token.

Create a workspace, mint an inference key, and call the same endpoint your production client will use. The console playground exercises the same execution path.

  1. 01
    Create a workspaceOwners and admins can create isolated workspaces inside an organization.
  2. 02
    Mint an inference keyScope the key to one workspace, with optional model and limit controls.
  3. 03
    Send an OpenAI-compatible requestUse your existing SDK or the curl example. The response is metered and settled.
Create a workspace Open the docs
QUESTIONS

Clear answers,
including the limits.

If a capability is not implemented, it should not appear as a feature. These are the boundaries customers and reviewers ask about first.

Is GateVia a model provider?

No. GateVia is a gateway and control plane. It routes requests to model endpoints, records usage and settlement, and gives teams a console for keys, catalog, finance and governance. The catalog currently includes models from multiple upstream authors through OpenRouter.

Which API shape is supported?

OpenAI chat-completions over /v1 and /api/v1, including server-sent events. Responses API and Anthropic-native skins are not implemented in this slice.

How does billing work?

Funding is prepaid and invoice-backed. The finance authority records credits and a ledger; the product requests settlement and never mutates a balance directly. External card/PSP processing is not activated.

Is automatic routing active?

No. Auto and Pareto policies remain shadow-only and default-off. They can be inspected, but live activation requires a separate owner authorization and an operator decision. The brake is a product feature, not a missing screen.

Can I bring my own provider key?

BYOK storage is implemented: provider secrets are encrypted at rest and obtained through an opaque per-attempt grant. OpenRouter is the current pilot provider. Other live provider onboarding remains a separate security and operations decision.

Is this production?

Not yet. This is an isolated private preview. Production is owner-authorized, but activation is gated on external ingress, paging, backup custody, KMS, production credentials and a zero-failure readiness run. We do not describe a preview as production.

What about SSO and enterprise identity?

SSO and SCIM are not implemented in this slice. The current identity model is account login, organizations, workspaces, invitations and role-based membership. Enterprise identity remains on the roadmap, not in the claim.

START HERE

Operate the model layer
with a real control plane.

Start with a workspace, inspect the catalog and send a request. Keep the keys, the price and the ledger in one place.

Create a workspace Read the docs Private preview · isolated staging · not a production SLA