Platform

Everything around the model, in one place.

Routing and failover, traffic health, budgets and limits, prompt versioning with an evaluation lab, caching, observability and usage records per customer — the operational work that sits between your product and whichever model answers. The model is the part you can swap, or supply yourself.

What the odnoga platform does

Multi-tenancy

A first-class tenant primitive everywhere — keys, budgets, analytics, prompts and audit logs are scoped per tenant.

Unlimited tenants per workspace
Programmatic provisioning via REST + SDK
Per-tenant Stripe customer & invoicing
Cascading limits: workspace → tenant → user

Key vault & rotation

Provider keys never leave the encrypted vault. Rotate, scope, revoke and audit without redeploying.

AES-256, envelope encryption
Scoping per tenant / model / region
Hot rotation with zero-downtime
Full audit trail of every key access

Budgets & rate limits

Soft caps warn users; hard caps stop the bleeding. Limits attach to any axis you care about.

Daily / monthly / lifetime budgets
RPM, TPM, concurrency throttles
Tier-based plans with overage rules
Webhooks on threshold breach

Prompt versioning

Treat prompts like code: branches, diffs, A/B tests, instant rollback. No more "what changed yesterday?".

Git-like version history with diffs
A/B routing by traffic percentage
Atomic production label switching
Per-tenant prompt overrides

Evaluation laboratory

Freeze one prompt version and its test cases, then compare the same workload across 2–8 active models.

Responses and assertion results side by side
Input/output tokens, billed cost and latency
Explicit one-shot capture of a real run
Review and redact before saving a case

Two-layer caching

Control deterministic odnoga answer reuse separately from vendor prompt caching. Each solves a different cost problem.

Separate opt-in controls
TTL, bypass and refresh controls
Creative calls can refuse answer reuse
Savings derived from request history

MCP server

Your customers connect odnoga into Claude Desktop, Cursor, Lovable or any MCP host — OAuth 2.1 with Dynamic Client Registration.

OAuth 2.1 + PKCE + DCR
Per-user scoped tokens
Tools exposed: chat, embeddings, image, search
Plug into Claude Desktop in seconds

Observability and traffic health

Every request is a row, and every caller is measured against its own normal. Trace cost and latency, then catch a stopped job or sudden surge.

Dashboards per tenant
Stopped and surged caller detection
CSV / Parquet export
OpenTelemetry traces

Smart routing

"auto" picks the best model for the request. Define fallback chains for cost, latency, or context length.

Cost-aware, latency-aware routing
Provider failover with retry policies
Per-tenant default model
Region pinning for compliance

Security & compliance

SOC 2 controls, EU-hosted option, PII redaction in logs, signed webhooks. Built for regulated SaaS.

SOC 2 Type II (in progress)
GDPR & DPA
PII redaction & log scrubbing
SSO (SAML, OIDC) on Enterprise — on request

Contextual support

Raise a ticket from the page where the issue happens, with useful page and workspace context attached for review.

Available on desktop and mobile
Ticket reference and status
Page context captured with the report
Open-ticket visibility for support staff

Architecture

How a request flows through odnoga

Your SaaS  ──►  odnoga Edge  ──►  Auth + Vault  ──►  Budget meter
                                                            │
                                                            ▼
                                                       Prompt registry
                                                            │
                                                            ▼
                                                      Router (auto/fallback)
                                                            │
                  ┌────────────────────────┬────────────────┼────────────────┬────────────────┐
                  ▼                        ▼                ▼                ▼                ▼
              OpenAI                 Anthropic           Gemini           Mistral          Groq …
                  │                        │                │                │                │
                  └────────────┬───────────┴────────────────┴────────────────┴────────────────┘
                               ▼
                       Observability + Billing  ──►  Stripe, Postgres, OTEL

Auth at edge

Bearer key + end-user header are validated before any work.

Stateless workers

Nothing is pinned to one machine, so capacity scales sideways.

Pluggable providers

A new provider is one adapter file — no change to routing, billing or the API you call.

Where the models come from

Two ways to reach a model. The same operations either way.

Routing, budgets, prompt versions, caching, traffic health and per-customer usage records do not depend on whose account pays the provider. Pick whichever suits you, and change your mind later without rewriting anything.

Use the odnoga catalogue

Every model in the catalogue, callable on day one with no provider accounts, no separate invoices and no minimum spend anywhere. You pay odnoga for what you use, at your plan’s rate.

Browse the catalogue

Bring your own provider keys

Already have accounts, negotiated rates or committed spend? Add your own OpenAI, Anthropic or Gemini keys and odnoga routes through them. The provider keeps billing you directly; everything on this page keeps working.

Included from Growth.

Integrations

Plays well with your stack

OpenAI SDK

Drop-in base URL replacement.

MCP hosts

Claude Desktop, Cursor, Lovable, Continue.

Stripe

Native billing, invoicing, webhooks.

Postgres / Supabase

All data in your stack-friendly format.

OpenTelemetry

Export traces to Datadog, Honeycomb, Grafana.

Honest limits

What odnoga is not

Not a self-hosted model runtime.

Bring your own provider; we route to it.

Not a vector database.

We integrate with pgvector, Pinecone, Weaviate; we do not store embeddings.

Not a model-training platform or a free-form AI playground.

The prompt workspace is built for controlled production operations: versioning, evaluations, A/B traffic and evidence-based promotion.