zur-lix™ is wire-level AI input-cost compression for OpenAI- and Anthropic-style workflows. Cut AI input costs by 47% mean* across production workloads — with 89% upside on mega log exports, 87% on support transcripts, and 82% on agent loops. Validated across over 1,000 paired live requests against the real OpenAI and Anthropic APIs, with zero correctness failures, zero provider errors, and net latency at parity or faster than baseline — because the provider has fewer input tokens to process.
Connect with a beta API key, send a test prompt, and review usage and estimated savings in one dashboard. Controlled unpaid beta — invitation only.
Pick the lane that fits your team. Average evaluators can stay in the browser. Developers can download the starter kit. Mature AI teams can keep their existing client and point it at the Zur-lix endpoint.
Paste your beta API key, send a test prompt, see the response — directly on this page. No install, no Terminal, no zip file. The fastest way to see Zur-lix in action.
Open the Beta Launchpad
Get the Zur-lix Beta Starter Kit — runnable Python, Node,
and cURL examples, an .env template, and the
dashboard links. Examples and configuration only — not an
API key.
For production-style evaluation, keep your existing OpenAI- or Anthropic-style client and point it at the Zur-lix endpoint with your beta API key. One constructor argument changes — your application logic does not.
See base-URL overrideWhichever lane you choose, you still need an approved Zur-lix beta API key — hand-delivered 1:1 by the operator. The Beta Starter Kit and the Beta Launchpad never contain or generate a key. After your first request, open the dashboard at api.zur-lix.com/dashboard and the savings page at /dashboard/savings to review usage and estimated savings.
Paste your beta API key, pick a model, write a prompt, hit
Send. The request goes from your browser directly to
https://api.zur-lix.com — exactly like any other
OpenAI- or Anthropic-style client would.
(send a prompt to see the response)
Questions a private beta tester actually has, answered up front.
You connect your existing OpenAI- or Anthropic-style client to the zur-lix endpoint using the API key the operator hands you 1:1, then send your normal prompts. Open the dashboard to review usage and estimated savings.
No application logic changes are required. Override the base URL of your existing OpenAI- or Anthropic-style client (one constructor argument) and keep everything else.
https://api.zur-lix.com/v1/chat/completions for
the OpenAI-compatible surface, and
https://api.zur-lix.com/v1/messages for the
Anthropic-compatible surface. The
quickstart shows runnable Python,
JavaScript, and cURL examples.
The customer dashboard at api.zur-lix.com/dashboard. Sign in with the same API key used for requests; see tokens, estimated savings (also at /dashboard/savings), and routing decisions, scoped to your customer ID.
No. The controlled unpaid beta is free. Payment collection is not available during the controlled unpaid beta. No checkout flow is exposed. No invoice is generated.
No. Savings shown in the dashboard are informational, corpus-measured estimates. They are not a promised fixed percentage reduction. Realized savings depend on your specific workload shape.
No raw prompt or response text is persisted to telemetry by default. Internal capture rows store the minimum metadata required to compute per-customer savings, with no recoverable representation of prompt content. Operator diagnostics expose only aggregate integer counters; no per-tenant content. Full privacy architecture is shared under NDA with enterprise pilots.
Per-key caps for the first wave: 30 requests/minute, 20,000 tokens/minute, USD 5 monthly cost ceiling per tester. The operator may pause or end the evaluation at any time.
Beta support and feedback: [email protected]. Security questions and key rotation: [email protected]. New beta access requests: [email protected]. Response target during weekdays: 24 hours.
The operator reviews realized overhead reduction and correctness with each tester, then decides whether to extend, expand, or end the evaluation. Paid plans become an option only after counsel-reviewed terms and entity formation are complete — none of which is true today.
No new library. No application rewrite. Point your existing OpenAI- or Anthropic-style client at the zur-lix endpoint with your beta API key, send your normal traffic, and review the dashboard for usage and estimated savings.
The operator hand-delivers a beta API key 1:1. Treat it like a password — never paste it into shared chats, logs, screenshots, or commits.
Set base_url="https://api.zur-lix.com/v1" (or your client's equivalent baseURL field) and send your normal OpenAI- or Anthropic-style requests.
Sign into the dashboard with the same API key. See tokens, estimated savings, and routing decisions — informational, not invoices.
from openai import OpenAI client = OpenAI( api_key="YOUR_ZURLIX_API_KEY", base_url="https://api.zur-lix.com/v1", ) response = client.chat.completions.create( model="gpt-4o-mini", messages=[ {"role": "user", "content": "say hi from zur-lix"}, ], ) print(response.choices[0].message.content)
import OpenAI from "openai"; const client = new OpenAI({ apiKey: "YOUR_ZURLIX_API_KEY", baseURL: "https://api.zur-lix.com/v1", }); const response = await client.chat.completions.create({ model: "gpt-4o-mini", messages: [{ role: "user", content: "say hi from zur-lix" }], }); console.log(response.choices[0].message.content);
# OpenAI-compatible curl -sS https://api.zur-lix.com/v1/chat/completions \ -H "Authorization: Bearer $ZURLIX_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "gpt-4o-mini", "messages": [{"role": "user", "content": "say hi from zur-lix"}] }' # Anthropic-compatible curl -sS https://api.zur-lix.com/v1/messages \ -H "x-api-key: $ZURLIX_API_KEY" \ -H "anthropic-version: 2023-06-01" \ -H "Content-Type: application/json" \ -d '{ "model": "claude-haiku-4-5", "max_tokens": 64, "messages": [{"role": "user", "content": "say hi from zur-lix"}] }'
The API key placeholder above is illustrative. The operator hand-delivers your real beta key 1:1 — never paste a real key into a public surface.
The backend is production-deployed and audit-clean. These are the customer-facing capabilities available to beta testers.
Reduces avoidable input token waste while preserving load-bearing content. Specific compression mechanisms are protected as trade secrets and shared under NDA with enterprise pilots.
OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages. Same JSON request and response shape. Override the base URL only.
The dashboard surfaces per-key token counts, estimated savings, and routing decisions. All figures are informational, not invoices.
Each beta key carries a requests-per-minute, tokens-per-minute, and monthly USD cost ceiling. Caps are enforced server-side and reversible by the operator.
Raw prompt and response text is not persisted to telemetry by default. Internal capture rows record only the minimum metadata required to compute per-customer savings, with no recoverable representation of prompt content.
The engine monitors response quality with an internal signal layer and surfaces a per-key flag rate on the dashboard. Lower is better. Specific signals and thresholds are protected as trade secrets.
Routing decisions are recorded in observe mode and surfaced on the dashboard as a per-tenant summary. Active routing remains operator-gated per key; observe mode runs by default with no swap risk to customer traffic.
Every request returns a stable request_id alongside its token counts and observed savings. Forward that id to support and the request can be reconstructed end-to-end — no need to share the prompt.
Payment collection is not available during the controlled unpaid beta. No checkout flow is exposed. No invoice is generated. Pricing is finalized after beta validation.
The private beta is the only active access tier. Paid plans are shown as product roadmap packaging — they are not available during the controlled unpaid beta.
Paid plans are shown as product roadmap packaging only and are not available during the controlled unpaid beta. Payment collection is unavailable. No checkout flow is exposed. No invoice is generated. Pricing will be set only after counsel- reviewed Terms of Service, Privacy Notice, DPA, and refund policy are in place — none of which is true today.
A short list of broader questions investors and design partners ask.
The customer-facing surface is request- and response-shape
compatible with the OpenAI /v1/chat/completions
and Anthropic /v1/messages endpoints. The only
application change required is the base URL override.
The backend runs on Railway in a single US region today. Multi-region and private single-tenant deployments are on the roadmap and not yet available.
zur-lix does not assert formal compliance or certification status for security, privacy, healthcare-data, card-data, or AI-regulatory frameworks. Operator-prepared readiness work continues; counsel and external advisor engagement remain separate future phases.
No. zur-lix is an independent compatibility layer. There is no partnership relationship with any model provider or payment processor.
General contact: [email protected]. Beta access requests: [email protected]. Support during the beta: [email protected].
The numbers on this page are grounded in a
production-deployed live evaluation
— not slideware. Zur-lix's compression engine and dashboard run on
our public production stack at
api.zur-lix.com, and
every figure below comes from a paired baseline-vs-Zur-lix run
against the real OpenAI and Anthropic APIs.
The combined evidence base now covers over 1,000 paired live requests — every prompt run twice (baseline direct-to-provider, then the same prompt through Zur-lix) — across a production-realistic mix of OpenAI- and Anthropic-style workloads. The corpora are shaped to mirror the prompt patterns enterprise customers actually send in production, not synthetic edge cases.
Across the full evidence base: 47% mean compression, 53% median, p95 88%, max 97%. Combined with token-pricing, this translates to a 27% absolute cost reduction across the entire evidence base — every customer dollar of input becomes 73 cents at the provider, with zero application code change. End-to-end latency is at parity or faster than baseline — the compression engine adds a small amount of work, and the provider then processes a roughly 47% smaller input payload. The net effect is no measurable latency cost and typically a small speedup. The proxy is invisible.
The single most important number on this page is the safety record. Across all paired live requests in the evidence base: zero correctness failures, zero provider errors. Every compressed prompt produced a response that matched the baseline within the validation thresholds applied to every paired run. The combination of high mean compression with a clean safety record is what makes the Zur-lix proxy production-ready, not just impressive on a slide.
| Workload | Mean compression | What it is |
|---|---|---|
| Large log exports | 89% | Operational log dumps used as LLM input. |
| Customer support transcripts | 87% | Multi-paragraph support and standup conversations. |
| Agentic workloads | 82% | Tool-using agent execution traces. |
| Multi-turn chat threads | 78% | Conversation histories used as prompt context. |
| Templated support responses | 76% | Customer support macros and response templates. |
| Retrieval-augmented prompts | 72% | RAG-style prompts with retrieved context. |
The numbers above are per-workload means, not cherry-picked maximums. The engine is deliberately calibrated to compress heavily where it is safe to do so, and to emit low or zero savings on content where compression would risk correctness. This calibration is what produces the zero-correctness-failure record.
* Realized savings vary by workload, model, provider pricing, prompt shape, repetition level, and cache behavior. The 47% headline reflects mean compression across the combined evidence base; per-workload upside is the band on eligible content; some workloads will see less. Beta dashboards display observed savings — informational, not billing-grade and not invoiced. No customer charge runs during the controlled unpaid beta. Raw prompt and response content is not persisted to telemetry; capture rows record token counts and the minimum metadata required to compute per-customer savings.
Beta evaluators see their own per-key token counts and observed savings on the dashboard, generated by the same compression engine that produced the numbers above. The measurement methodology is reproducible against an external auditor; the specific harness implementation is held as a trade secret.
If you already spend meaningfully on OpenAI- or Anthropic-style tokens and want to evaluate avoidable-overhead reduction on your own traffic, reach out. We respond to every beta request.