What is ScaiGrid

ScaiGrid is a unified AI gateway. It sits between your application and every large language model provider you might want to call — OpenAI, Anthropic, Azure, Mistral, Google, Qwen, your own self-hosted models, anything else — and exposes a single, stable HTTP API on top of all of them.

You integrate once. Swap models, add providers, change routing policies, enforce budgets, and add capabilities without touching your application code.

What it gives you#

One endpoint for every model. Your application calls POST /v1/inference/chat. ScaiGrid handles provider selection, authentication to the upstream, retries, failover, and response normalization. Switching from GPT-4 to Claude Opus is an admin-UI toggle, not a code change.

Multi-tenancy, from day one. ScaiGrid has a three-level tenancy model: Partners, Tenants, Users. Every request is tagged with its tenant and user. Usage, budgets, rate limits, and access controls all scope cleanly. If you're a platform serving multiple customers, you don't need a second layer.

Accounting, budgets, and audit. Every completion is counted. Spend is tracked per tenant, per user, per model. Budgets enforce themselves at the gateway — a customer who hits their limit gets a 429, you never get a surprise bill. Full audit log for security review and compliance.

Pluggable modules. ScaiGrid is the platform, not the product. Modules add capabilities: an embeddable chatbot widget (ScaiBot), vector-and-graph knowledge search (ScaiMatrix), typed message queues (ScaiQueue), sandboxed code execution (ScaiBunker), and more. Modules register themselves at startup, get their own URL namespace, their own permissions, their own admin UI pages.

Drop-in OpenAI compatibility. An /oai/v1/ surface speaks the OpenAI API dialect exactly. Existing code that points at api.openai.com/v1/chat/completions can switch to ScaiGrid by changing the base URL. See OpenAI Compatibility. For new code, use the native /v1/ API — it's richer.

What it's not#

• Not a model. ScaiGrid doesn't train or host models itself. It routes to providers that do.
• Not a notebook or playground. There's an admin UI for operators and a chat playground for experimentation, but ScaiGrid is primarily an API.
• Not a billing system. ScaiGrid tracks usage and enforces budgets; it doesn't generate invoices or charge credit cards. Surface the usage data to your own billing pipeline.

Who it's for#

• Platforms building on top of LLMs that want a stable integration surface and multi-tenant cost control.
• Teams integrating multiple providers who don't want to maintain N provider-specific clients.
• Operators of self-hosted models who want a front door with auth, rate-limiting, and observability.
• Anyone migrating off the OpenAI SDK but not willing to rewrite their inference code. Drop-in works.

What's next#

• Philosophy — the design principles behind the API shape.
• Architecture — how the pieces fit together.
• Quickstart — your first chat completion in five minutes.