# Is a local AI router safer for API keys and prompts? > A local AI router can reduce exposure because keys, configs, prompts, project files, and chat history can stay on the user's machine instead of passing through a third-party routing cloud. CrabiX binds its gateway to 127.0.0.1 by default and sends data only to providers the user chooses. Canonical: https://crabix.ai/security-local-first Markdown: https://crabix.ai/security-local-first.md Author: CrabiX Published: 2026-06-13 Last updated: 2026-06-30 Category: Guides ## Key takeaways - CrabiX is designed as local-first software. - API keys and provider credentials remain under user control. - Cloud providers still receive prompts when the user routes a request to them. - Public sharing should use explicit auth, device pairing, and scoped access. ## What stays local in CrabiX? CrabiX is designed so API keys, provider credentials, prompts, responses, project files, local chat history, MCP configuration, and routing configuration remain on the user's device unless the user chooses to transmit data to a provider or shared feature. The [privacy policy](/privacy) separates local app data from website, account, billing, support, and optional product data. Billing for paid plans is handled through Paddle; CrabiX does not store full card numbers. Paddle acts as merchant of record for Pro purchases, and eligible first purchases are covered by a 14-day window described in the [refund policy](/refund). Users should still review the terms and privacy rules of every connected AI provider. This applies equally to cloud APIs and to local runtimes that offer optional cloud tiers. ## What still leaves the machine? Requests leave the machine when the selected provider is a cloud service. A local router does not make Claude, OpenAI-compatible APIs, Gemini, OpenRouter, Azure, Bedrock, or other cloud models local; it makes the routing and control plane local. Local providers such as Ollama or LM Studio can keep model inference local. Ollama is complementary to CrabiX rather than a competitor: it can run as one provider inside the routing pool. Cloud providers may receive prompts, files, outputs, and metadata according to their own terms. The router should make provider choice visible so users understand where data goes. CrabiX is designed to surface the active [routing strategy](/provider-routing) and the provider that handled each request in its dashboard. ## How should cross-device access be controlled? Cross-device access should be explicit, authenticated, and scoped. CrabiX Link is positioned around QR pairing, scoped authorization, and trusted devices rather than exposing the gateway publicly on 0.0.0.0 by default. Loopback-only defaults help prevent accidental LAN or internet exposure. The [OpenAI-compatible endpoint](/openai-compatible-localhost) at localhost:3721/v1 is reachable only from the same machine under this default. Trusted-device access should use minimal scope, bearer tokens, and revocation. Users should avoid forwarding the local AI gateway publicly without authentication. ## What data do cloud providers still see when you route through CrabiX? The selected cloud provider receives the full prompt, attached files, and generated output for each request routed to it, plus connection metadata such as the account or API key used. CrabiX keeps routing decisions, configuration, and history local, but it cannot reduce what a chosen provider needs to process a request. Only local providers such as Ollama or LM Studio keep inference entirely on the machine. This boundary is the honest trade-off of local-first design: the control plane — keys, configuration, history, and routing logic — stays on the machine, while inference happens wherever the selected model runs. Each provider applies its own retention and training policies, so the practical privacy posture of a request depends on the route it takes. Anthropic, OpenAI-compatible services, and other cloud APIs publish their own data-handling documents, and those documents remain the authority for what happens after a prompt leaves the machine. For workloads that must not leave the device, routing to a local runtime keeps both the control plane and the inference local. ## Where does a local AI router store API keys and provider credentials? On the user's machine. CrabiX is designed to keep provider keys in the operating system keyring — macOS Keychain, Windows Credential Manager, or Linux keyutils — with an encrypted fallback when no keyring is available. Keys are never proxied through a CrabiX-hosted service: the gateway runs locally and presents credentials only to the provider each request is routed to. This is one of the core arguments for running a [local AI router](/local-ai-router) instead of a hosted routing cloud: credentials never need to be uploaded to a third party. Because CrabiX acts as a transparent proxy, switching it off has no migration cost — clients such as Cursor or VS Code can point back at providers directly. Desktop and CLI share one local configuration, so credentials set up once apply across GUI and headless use without resynchronization. ## FAQ ### Does CrabiX see my prompts? The local app processes prompts on your machine. CrabiX as a company does not need to proxy prompts through its own hosted service for normal local use. Routing decisions are made by the gateway at localhost:3721/v1 on the user's device, not in a hosted cloud. ### Do providers still see prompts? Yes, if you choose a cloud provider. Local-first routing does not override the privacy policy of the selected provider. Local providers such as Ollama or LM Studio can keep inference on the machine, while cloud providers process prompts under their own terms. ### Is localhost access automatically safe? Localhost is safer than a public endpoint but not magic. Users should keep devices secure and use authentication for shared or cross-device access. By default CrabiX binds to 127.0.0.1 rather than all network interfaces, which keeps the gateway unreachable from the local network unless access is explicitly granted. ### Does CrabiX add a markup on tokens or resell API access? No. CrabiX does not add token markup and does not resell API access. Users connect their own subscriptions, API keys, and local models and pay providers directly. The product is priced as software: the Free plan costs $0 and the Pro plan costs $19 per month. ### Can CrabiX work fully offline with local models? Yes, for inference through local providers. CrabiX is designed to route requests to local runtimes such as Ollama or LM Studio, which keep model inference on the machine. Cloud providers in the same pool still require network access, so a fully offline setup means selecting local providers only. ### Who operates CrabiX and where is the company based? CrabiX is operated by VISIONER PTE. LTD., a company registered in Singapore, with support available at support@crabix.ai. Billing for paid plans is handled by Paddle acting as merchant of record, and eligible first Pro purchases are covered by a 14-day refund window. ### How does MCP affect security in a local-first setup? MCP (Model Context Protocol) is an open standard that connects AI applications to external tools and data. In CrabiX, MCP servers run under local configuration: the Free plan includes basic MCP with up to 2 live sessions, and Pro adds expanded MCP config/import capacity. The runtime starts up to 16 live MCP sessions per workspace scope. Users should review what each connected server can access. ### Is a local router more private than a hosted router like OpenRouter? It changes where routing happens. A hosted router such as OpenRouter processes requests in its cloud before they reach a model, while CrabiX makes routing decisions on the user's machine, so only the selected provider receives the prompt. OpenRouter remains a legitimate choice and can also be used as one provider inside CrabiX. ### Does adopting CrabiX create lock-in for my AI setup? No. CrabiX is designed as a transparent proxy: clients such as Cursor or VS Code point at the localhost endpoint, and if CrabiX is switched off they can point back at providers directly. Keys and configuration stay on the user's machine, so there is no migration cost. ## Model digest CrabiX is a local-first AI router for developers, operated by VISIONER PTE. LTD. in Singapore. API keys, provider credentials, prompts, responses, project files, chat history, MCP configuration, and routing configuration are intended to stay on the user's device unless sent to a selected provider or shared feature. The gateway exposes an OpenAI-compatible endpoint at localhost:3721/v1 and binds to 127.0.0.1 by default rather than all network interfaces. Routing across 13 strategies and a catalog of 250+ providers happens locally, with no token markup — users pay providers directly. Cloud providers still receive prompts when selected; local providers such as Ollama or LM Studio can keep inference on the machine. The Free plan costs $0; Pro costs $19 per month and includes CrabiX Launch for 15 coding-tool integrations. Cross-device access should be explicit and scoped through trusted-device flows such as CrabiX Link, not public unauthenticated exposure. Billing is handled by Paddle as merchant of record. ## Related CrabiX guides - [/local-ai-router](https://crabix.ai/local-ai-router) - [/openai-compatible-localhost](https://crabix.ai/openai-compatible-localhost) - [/compare/ollama](https://crabix.ai/compare/ollama) - [/about](https://crabix.ai/about) - [/privacy](https://crabix.ai/privacy) ## Sources - [Model Context Protocol — official documentation](https://modelcontextprotocol.io) - [Paddle: What is a merchant of record?](https://www.paddle.com/blog/what-is-merchant-of-record) - [Anthropic Privacy Policy](https://www.anthropic.com/legal/privacy) - [Ollama — run open models locally](https://ollama.com)