CEDX AI Gateway · AI
CEDX AI Gateway sits between your apps and the models: a virtual key per team, routes with primaries and counted fallbacks, caching, rate limits — and a ledger that writes latency, tokens and cost on every single request.
The most expensive request in the demo window cost $30.25 — route.support.chat.27 on assist-large, 8.5K tokens in. The cheapest visible cost is $0.0000. Both are rows in the same table, which is the entire point of a gateway.
Runs on demo data. Northline is the software's demo workspace, not a customer.
What it is
340 requests in the sample, each with route, key, model, status, latency, tokens in and cost. route.support.rank.22 answered in 16ms for $0.0000 on assist-fast; route.support.chat.27 took 2.8 seconds and $30.25 on assist-large. Same table, same columns — no arguing about where the money went.
36 routes, each naming a team, a primary model and a fallback count. route.support.v3 runs assist-balanced with 2 fallbacks over 48,200 hits; route.onboard.batch.34 shows 14.4% of its traffic falling back and is flagged Error — which is exactly when you want a gateway and not a wrapper.
The usage grid keeps 180 day-key rows — 176,843 requests and $95K of usage-grid spend — and the sample's spend share is blunt: assist-reasoning takes 52% ($788.21), assist-large 29%. The two cheapest model families share 3% between them.
Product tour
Not a mockup and not a concept deck. This is what opens at /app/ai-gateway.
01 — Overview
340 sampled requests: 265 ok, 45 cached, 27 rate-limited, 2 error, 1 timeout. The health ring reads 0.9% errors and 7.9% 429s; 9 of 48 virtual keys carry critical risk. The alerts arrive with the fix attached: vk-code-review at 94% of its RPM soft-cap — raise the cap to 300 or route the canary off peak.
02 — Requests
The request ledger is the ground truth: req id, route, virtual key, model, status, latency, tokens in and cost, with a Sessions tab beside it. The $30.25 request and the $0.0000 request are twenty rows apart.
03 — Usage & cost
Spend and cache analytics on one screen: $1.5K of sample request spend against a $95K usage grid, 1282K tokens in and out, and a spend-share panel that puts 52% of the money on assist-reasoning. The day-key table rolls it up with 429s counted per row.
04 — Routes
Thirty-six routes with team, primary model, fallback count, hits, fallback share, error share, latency and spend. route.batch.jobs carries 62,000 hits on assist-fast at a 6.8% error share — Warning, visible, and nobody had to grep a log to find it.
Who runs it
Roles, not references. We have no named customers yet, so nobody in these photographs is quoted, credited or claimed as one.
Owns routes and providers: sets primaries and fallbacks, watches fallback share, and answers for the 10 routes flagged degraded.
36 routes · 10 degraded
Reads the day-key grid the way finance reads a ledger. 52% of sample spend on one model is a conversation started early, not a surprise at month end.
$95K · 180 day-key rows
Gets the alert when vk-code-review hits 94% of its RPM soft-cap, and chooses between raising the cap and routing the canary off peak.
429 rate · 7.9%The shape of it
Every figure below is legible in the captures above. Nothing here is a projection of your estate — it is the state of the demo data.
How it runs
Each team or workload gets a virtual key — 48 in the demo — so usage, limits and risk attach to something ownable.
The request enters on a route with a primary model and counted fallbacks; route.support.v3 carries 2 over 48,200 hits.
Soft-caps and rate limits bite before budgets do: vk-code-review at 94% RPM raised an alert with two fixes attached.
Every request writes its tokens and cost to the ledger; the usage grid rolls it up by day and key, 429s included.
One record
The same model lineup — assist-mini through assist-reasoning, vision-edge — appears on every AI screen in the estate, because the calls share one ledger.
The models on Copilot's prompt registry are the models these routes carry — which is why its cost-per-run column and this ledger speak the same language.
CEDX AgentsEvery fleet row's model column is a routed call. The agents' $0.002-to-$0.37 cost-per-run spread is this routing, seen from the other side.
CEDX SpeechSpeech meters its listen and speak models per job and per session; per-call cost and rate discipline is what a gateway is for.
CEDX KnowledgeKnowledge's AI answer is synthesized from retrieved cards — a model call that can be routed, cached and priced like any other.
Limits
Finding this out on the third call is worse for you than reading it here, and worse for us.
Start
Open it right now
Your keys, your routes
The gateway with the rest of it
Questions
Yes. Every screenshot is a capture of the running build and you can open the same build at /app/ai-gateway. It runs on demo data — Northline is the demo workspace, not a customer.
The route falls back, and the fallback is counted. Routes carry a primary and an FB count — route.support.v3 fired its fallback in the demo window and the alert's advice was to keep the weights and watch provider health, not to panic. Fallback share is a column on every route, up to 15.6% in view.
Three ways, all on screen: every request writes its cost to the ledger; soft-caps alert before budgets blow — vk-code-review was caught at 94% of its RPM cap; and caching took 13.2% of the sample after the TTL change.
That is what virtual keys are for. Each of the 48 keys in the demo carries its own usage, limits and risk flag — 9 are flagged critical risk — and the usage grid breaks spend down by day and key.
No certification has been issued. What we can evidence about hosting, encryption, tenant isolation and production access is written up on the security page.
Live build, demo data, no card. Then check what your most expensive request would have cost.