llmcloud.ai
For providers

Put your endpoints in the router pool.

llmcloud.ai routes millions of requests across upstream providers. If your endpoints win on price, latency, or quality, we send you traffic and pay commissions monthly.

Developer distribution
Instant reach into every app using our SDK — no per-app integrations.
Fair routing
Requests are matched on capability, price, and live health. No pay-to-win.
Zero rewrite
If you speak OpenAI-compatible /v1/chat/completions, you're ready.

Requirements

OpenAI-compatible API
/v1/chat/completions, /v1/embeddings, streaming, and tool calls.
Public health endpoint
GET /health returning region, model list, and version.
≥99.5% 30-day uptime
Measured from our probes across 3 regions.
Rate-limit headers
Standard X-RateLimit-* headers so we can back off cleanly.
Per-model pricing manifest
JSON manifest we can poll for input/output/image pricing.
Signed webhooks
Optional but recommended for usage reconciliation.

How onboarding works

  1. 01
    Apply
    Fill out the form below. We reply within 3 business days.
  2. 02
    Sandbox
    You get a scoped test key. We run compatibility, streaming, and tool-use suites.
  3. 03
    Load & latency
    We benchmark p50/p95 tps and TTFT across 3 regions for 48h.
  4. 04
    Publish
    Approved endpoints appear in the catalog and enter the router pool.
  5. 05
    Settle
    Monthly reports plus programmatic access to per-model usage and commissions.

Certification timeline

Nothing is admitted on a self-report. Each stage produces measurements we publish on your public provider page, so developers can see exactly why you carry the grade you carry. The full rubric and its thresholds are documented in provider standards.

StageTypical durationWhat we measureFails if
Application review3 business daysCompany details, endpoint inventory, pricing manifest, security contact.No published pricing, or no named security contact.
Compatibility suite1–2 daysChat, streaming deltas, tool calls, JSON mode, embeddings, error shapes against a 400-case suite.Any tool-call payload that does not parse, or non-standard SSE framing.
Load and latency burn-in48 hoursp50/p95 time to first token and sustained tokens per second from probes in three regions.p95 TTFT above 3 s, or throughput collapse beyond 40% at rated concurrency.
Weight verification1 dayGolden-prompt fingerprinting against reference weights to detect quantization or substitution.Served precision differs from what the manifest declares.
Publishsame dayProvider page, published grade, and entry into the router pool at your list price.
Continuous re-certificationongoingRolling 30-day uptime, drift probes, and price-manifest polling.Grade drops automatically; sustained failure removes you from the pool.

Weight integrity rules

The single most common complaint developers have about multi-provider gateways is silent substitution: a quantized build, a shorter context window, or a different fine-tune served under a name that implies the original weights. Providers in the llmcloud pool agree to four rules, and we probe for violations continuously rather than trusting a checkbox.

Declare served precision
The manifest states the exact precision and build served per model. FP8 sold as FP16 is grounds for removal, not a warning.
No silent substitution
Under load you may shed traffic with a 429. You may not swap in a smaller or different model to absorb it.
Publish the real context window
The advertised window must be the window that actually works, not the one the architecture theoretically supports before your serving config truncates it.
Announce changes before they ship
Precision, quantization, context, or price changes require notice so the catalog and routing weights update before developers are surprised.

Provider FAQ

What does it cost to list on llmcloud?

Nothing. There is no listing fee, no placement fee, and no way to pay for better routing position. The gateway also takes no cut of your token revenue — developers pay your list price directly, which is why routing position can only be earned through price, latency, and measured quality.

How is traffic allocated between providers serving the same model?

By score, not by contract. Each provider/model pair is scored on rolling success rate, quality rank, measured time to first token, blended price, and remaining capacity headroom. Winners get more traffic and losers decay out, with the exact weights published on the routing page.

Does llmcloud's own hosted fleet get preference?

No. llmcloud-hosted endpoints pass the same certification stages, publish the same grade, and are scored by the same function as any third party. When an external provider is cheaper or faster, the router sends the request there.

What happens when our uptime slips?

Grades are recomputed continuously from our own probes, not from your status page. A dip lowers your score and reduces traffic automatically; sustained failure below the 99.5% 30-day threshold removes you from the pool until a re-certification burn-in passes.

Do we have to support every endpoint?

No. Many providers list chat completions only. The catalog records exactly which endpoints, modalities, and features each provider supports, and the router only sends requests a provider can actually serve.

Can we serve region-restricted capacity only?

Yes. Declare the jurisdictions you serve and the router will only send requests pinned to those regions. This is how residency-constrained and sovereign traffic is handled.

Apply

Fields below build your provider profile draft. Nothing goes live until you review it.

By submitting you agree to our provider terms and data-processing addendum.