AI orchestration layer · self-hosted
Every AI model.
One orchestration layer.
Deploy once. Route anywhere. Scale without provider lock-in.
REQUEST RESPONSE RETRY / FAILOVER
The problem
Your infrastructure shouldn't depend on one model.
A single provider is a single point of failure. One rate limit, one outage, one deprecation — and your product is offline. Orchestration makes the failure somebody else's problem: the router's.
WITHOUT NEXORA
200 OKWITH NEXORA
200 OK · ALWAYSIntelligent routing
A router that understands providers, not just URLs.
Every request is scored against live health, latency, cost and rate-limit state — then dispatched to the best available route. Automatically. On every call.
Model selection
Resolve a model name to the best live deployment.
Latency optimization
Prefer the fastest healthy route, per region.
Cost optimization
Route to the cheapest provider that meets the SLA.
Fallback chains
Ordered candidates for every model, tried in sequence.
Provider health
Continuous scoring from real request outcomes.
Weighted routing
Split traffic across providers by explicit weight.
Automatic retries
Typed error detection with Retry-After-aware backoff.
Circuit breakers
Failing routes cool down and re-enter gradually.
Live provider status
Account-level state, visible and queryable.
ROUTING TABLE · LIVE▍
policy: weighted · health-aware
DECISION LOG
→ resolve claude-sonnet-5 · 3 candidates · policy: latency
→ anthropic/acct-1 healthy · p50 455ms · dispatch
✓ 200 · streamed 1,204 tokens · $0.0038
Multi-provider
Every ecosystem. One integration.
Connect the providers you already use. Nexora normalizes authentication, streaming, errors and rate limits behind a single interface.
OpenAI
GPT models with pooled keys, org-aware limits and automatic account failover.
Anthropic
Claude models with native streaming, extended thinking and account rotation.
Google Gemini
Gemini family, plus Vertex AI deployments for enterprise workloads.
Groq
Ultra-low-latency inference — the router's favorite for speed-critical paths.
OpenRouter
Hundreds of models behind one key. A natural universal fallback tier.
Ollama
Keep sensitive workloads on hardware you own. Zero external calls.
LM Studio
Local OpenAI-compatible server, treated as a first-class routing target.
Azure OpenAI
Enterprise deployments with regional routing, quotas and private networking.
AWS Bedrock
Bedrock models with SigV4 auth and per-region failover built in.
Custom Providers
Any OpenAI-compatible endpoint becomes a routed provider in minutes.
46+ PROVIDERS BUILT IN · OAUTH FOR CLAUDE, GEMINI & CODEX · EVERY PROVIDER SUPPORTS ACCOUNT POOLS
Multi-account scaling
Rate limits are a per-account problem. You have a pool.
Connect multiple authenticated accounts per provider and Nexora treats them as one capacity pool. When an account reaches its limit, traffic shifts to the next before a single request fails — automatic capacity management, not manual key juggling.
Per-account health. Every account carries its own live state, error counts and cooldown window.
Retry-After aware. Rate-limit responses are parsed, not guessed — accounts re-enter exactly when the provider says so.
Circuit breakers. Repeated failures open the breaker; recovery is gradual and observed.
OPENAI · PRODUCTION POOL
4 accounts
POOL THROUGHPUT
steady · 0 dropped requests
Architecture
A request's journey through Nexora.
From client call to model response: every layer is observable, retryable and replaceable. This is the whole system — no hidden magic.
Client
SDKs · REST · WebSocket · SSE
Nexora Gateway
auth · rate control · streaming
Agent Runtime
turn engine · sub-agents · skills
Router
weighted · health-aware · cost + latency
Memory
semantic memory · RAG · pgvector
Tool Execution
90+ builtin tools · permission gate
Provider Layer
46+ providers · account pools · breakers
LLMs
every model · one interface
Platform
Everything an AI stack needs between app and model.
Orchestration, routing and reliability primitives that would otherwise live in your codebase — maintained, tested and self-hosted.
Agent orchestration
Compose agents, sub-agents and skills into one coordinated runtime.
Provider routing
Health-aware, weighted routing across every connected provider.
Multi-account management
Pool authenticated accounts per provider for automatic capacity management.
Streaming responses
Token-level streaming over WebSocket and SSE, provider-agnostic.
Observability
Latency, cost and health metrics for every request, account and route.
Retry logic
Typed rate-limit detection with Retry-After-aware backoff.
Circuit breakers
Failing providers cool down automatically and re-enter when healthy.
Tool execution
90+ builtin tools behind a permission gate on every call.
Model abstraction
One request shape for every model — swap providers without code changes.
OpenAI-compatible API
Point any existing SDK at Nexora and it just works.
Custom providers
Register any OpenAI-compatible endpoint as a first-class provider.
Self-hosting
Runs on your infrastructure. Your keys, your data, your network.
Provider chains
Define the chain once. Nexora runs the failover.
Order your provider accounts and Nexora does the rest: it resolves the model, retries rate limits, fails over on errors and streams the response. No handling code in your app.
# chain · gpt-5 production
model: gpt-5 # or any connected model
route:
- openai/acct-1 # primary
- openai/acct-2 # failover on 429
- azure/gpt-5 # failover on outage
- openrouter # last resort
CONFIGURED IN THE CONSOLE OR THE CLI — NO CODE IN YOUR APP
→ resolve gpt-5 · candidates: openai/acct-1, azure/gpt-5, openrouter
→ openai/acct-1 · 429 rate limited · retry-after 18s
→ failover openai/acct-2 · healthy · p50 412ms
✓ 200 · streamed 812 tokens · $0.0041 · 1.2s total
A rate limit occurred, an account failed over, and the request succeeded — zero lines of handling code.
- 0M+
- requests routed
- 0.00%
- provider uptime
- 0+
- supported models
- 0+
- connected providers
- <0ms
- avg routing overhead
Clients
One instance, three ways in
Run the platform however you work. Every client talks to the same self-hosted Nexora instance.
Pricing
Simple, honest pricing
The open-source core is free forever. NexoraCloud adds what your team needs to run agents in production. You pay for the service and support, not for a crippled product.
The full Nexora platform, self-hosted on your own infrastructure.
- Full agent builder & orchestration
- 46+ LLM providers
- 90+ built-in tools & skills
- RAG & semantic memory
- CLI + mobile app
- Community support (GitHub)
For small teams getting started with production agents.
- Everything in Community
- Up to 5 agents
- Up to 3 users
- License key + verified downloads
- Usage dashboard
- Email support
For growing teams running agents across real workflows.
- Everything in Starter
- Unlimited agents
- Up to 10 users
- Marketplace access
- Usage analytics & cost breakdown
- Cost alerts & budget limits
- Priority support
For teams that need audit trails and compliance.
- Everything in Pro
- Up to 50 users
- Full audit logs
- GDPR data export
- SOC2 data retention
- IP allowlist & granular RBAC
- White-label branding
For organisations needing SSO, provisioning, and a contract.
- Everything in Team
- Unlimited users
- SAML 2.0 SSO + SCIM provisioning
- OIDC support
- Dedicated support + SLA
- Custom contract & invoicing
Build AI without infrastructure limits.
Self-hosted, open core, every provider. The orchestration layer is the only part of your AI stack you should never have to think about.
SELF-HOSTED · MIT OPEN-SOURCE CORE · YOUR INFRASTRUCTURE