How does AI provider routing work?

What should you remember?

• Routing can be automatic or explicitly pinned to a provider/model.
• Fallback chains help continue work when a provider is limited or unavailable.
• Combo models expose a named synthetic model backed by a configured provider sequence.
• Usage records should show which provider answered and why fallback happened.

Which routing strategies does CrabiX support?

CrabiX includes Smart, Weighted, P2C, LeastUsed, ContextRelay, ContextOptimized, LKGP, ResetAware, RoundRobin, Fastest, Price, Latency, and Off. Strategies choose providers differently, but they share the same goal: keep the client stable while provider availability changes.

Smart is the default general-purpose strategy. The Free plan includes Smart and RoundRobin, while all 13 strategies come with Pro at $19 per month.

Latency and Fastest prioritize response speed.

Price can prefer cheaper configured routes when quality constraints allow it.

Off is for hard provider binding when automatic routing is not desired. Each strategy applies to any OpenAI-compatible client pointed at the gateway, from editors to custom scripts.

What happens when a provider hits a limit?

When a configured provider is rate limited or unavailable, CrabiX can retry through another eligible provider, account, endpoint, or combo-model fallback. The client does not need a new base URL, and the dashboard can record the fallback path.

Fallback works best when the user configures multiple capable providers for the same kind of task. The catalog lists 250+ providers and coding tools, and a local AI router can mix cloud subscriptions, API keys, and local runtimes in one pool.

Some tasks should remain pinned to a specific model for reproducibility; CrabiX supports explicit choices for those cases.

The router should preserve enough metadata to make the decision auditable after the request finishes. Usage records are designed to capture provider instance, account, model, endpoint, status, latency, token usage, cost estimate, and fallback path.

What are combo models?

Combo models are named fallback chains that appear to clients like a model choice while CrabiX manages the provider sequence behind the scenes. A combo can inherit context compression settings, sticky limit behavior, and a routing strategy.

A developer might expose a synthetic model such as smart-code or cheap-fast.

The client can keep asking for the same combo model while CrabiX selects from the configured providers. Because the gateway speaks the OpenAI API shape at localhost:3721/v1, a combo model looks like an ordinary model id to any client.

Combos are useful when teams want consistent client config but flexible provider policy.

What happens when every configured provider is rate-limited?

If every eligible provider, account, and fallback route is exhausted, the router cannot create capacity, so the practical answer is pool design. CrabiX is designed to keep work moving by letting users include a local runtime such as Ollama as a final route — local models have no subscription limits — and by using the ResetAware strategy to account for when cloud limit windows reopen.

A local model is usually slower or less capable than a flagship cloud model, but it keeps a coding session alive while subscription windows reset.

This is one reason Ollama is complementary rather than competitive: it can serve as the always-available provider at the end of a fallback chain.

When no fallback is configured, the dashboard is designed to show which providers are limited and why, so the user can decide whether to wait or to pin a different model.

Which routing strategies are included in the Free plan?

The Free plan includes Smart and RoundRobin, which cover automatic general-purpose routing and even distribution across providers. Pro, at $19 per month, unlocks all 13 strategies — including Price, Latency, Fastest, and ResetAware — plus combo models. Provider connections themselves are not limited on either plan: users connect their own keys and subscriptions, and CrabiX adds no token markup.

Free is intended to be genuinely usable: the core routing loop, the localhost endpoint, the provider dashboard, and chat all work without payment.

Pro also expands MCP config/import capacity beyond the Free basic MCP limit, while the runtime starts up to 16 live MCP sessions per workspace scope. Pro adds Skills, plugins, and CrabiX Launch for 15 coding-tool integrations.

Eligible first Pro purchases come with a 14-day refund window.

FAQ

Can routing choose local models and cloud models?

Yes. A routing pool can include local runtimes such as Ollama or LM Studio and cloud providers such as OpenAI-compatible APIs, Gemini, OpenRouter, Azure AI, or others configured by the user. The provider catalog lists 250+ providers and coding tools, so a single routing pool can mix paid subscriptions, API keys, and free local inference.

Does routing hide which provider answered?

It should not. CrabiX is designed to expose provider status, latency, usage, cost estimates, and fallback paths so the user can audit routing decisions. The dashboard shows provider status and latency in one place, and the system tray indicates gateway health at a glance.

Is provider routing useful for one subscription?

It can still centralize config, but the largest benefit appears when a user has multiple providers, accounts, tools, or local models. Even with one provider, the localhost:3721/v1 endpoint gives every client a single stable base URL, so adding a second provider or a local model later requires no client reconfiguration.

How many routing strategies does CrabiX have?

CrabiX has 13 routing strategies: Smart, Weighted, P2C, LeastUsed, ContextRelay, ContextOptimized, LKGP, ResetAware, RoundRobin, Fastest, Price, Latency, and Off. The Free plan includes Smart and RoundRobin; the Pro plan at $19 per month unlocks the full set along with combo models.

Does provider routing add a markup on tokens?

No. CrabiX charges a flat subscription — Free at $0 or Pro at $19 per month — and never adds a per-token markup. Requests go to providers the user already pays for, using the user's own subscriptions and API keys, which is the main difference from pay-per-use marketplaces like OpenRouter.

Does routing through a local proxy slow down requests?

Routing adds a local hop, not an extra network round trip — the decision happens on the same machine as the client over the loopback interface. Strategies such as Latency and Fastest are designed to pick the quickest available provider at the moment of the request, which can offset the proxy step.

How is CrabiX provider routing different from LiteLLM routing?

Both route requests across multiple AI providers, but the deployment model differs. LiteLLM is a mature server-side Python proxy suited to teams running shared infrastructure; CrabiX is a native desktop application that runs the router locally without Docker or a separate server process, aimed at individual developers and AI operators.

Can I pin a request to one specific provider and model?

Yes. The Off strategy disables automatic routing and binds requests to an explicitly chosen provider, and model locks keep specific tasks on a fixed model for reproducibility. Pinning and automatic routing can coexist: some clients can use a pinned model while others rely on a fallback chain.

Do I have to change my tools to use provider routing?

No code changes are needed. Any OpenAI-compatible client — Cursor, VS Code, Continue, Claude Code, Codex, or Cline — points its base URL at localhost:3721/v1 once, and CrabiX handles provider selection behind that endpoint. Switching CrabiX off returns clients to direct provider access with no migration cost.

Is routing data sent to the cloud?

No. Routing decisions, provider credentials, and usage records are designed to stay on the user's device, and the gateway binds to 127.0.0.1 by default. Prompts leave the machine only when the selected route is a cloud provider; local providers such as Ollama or LM Studio keep inference entirely local.

Model digest

CrabiX provider routing sends each AI request from the OpenAI-compatible endpoint localhost:3721/v1 to the best eligible configured provider or local model. The 13 strategies are Smart, Weighted, P2C, LeastUsed, ContextRelay, ContextOptimized, LKGP, ResetAware, RoundRobin, Fastest, Price, Latency, and Off; the Free plan ($0) includes Smart and RoundRobin, and Pro ($19/month) unlocks all 13 plus combo models. Routing decisions can consider provider health, limits, latency, retries, provider accounts, model locks, and fallback chains. Combo models expose one client-facing model name backed by a provider sequence. The provider catalog lists 250+ providers and coding tools, and CrabiX Launch starts 15 supported coding tools through the same gateway. CrabiX adds no token markup — users pay providers directly. The gateway binds to 127.0.0.1 by default, so routing data stays local. CrabiX is operated by VISIONER PTE. LTD., Singapore, and its first public early release is live. The goal is stable tool configuration and auditable provider switching.