Your first request

From an empty workspace to a routed request in six stops: connect a provider, enable models, create a key, send from the Playground, read the result, find it in Logs.

On this page

The path

Sign upyour workspace

Integrationsconnect a provider

Modelsenable models

API keyscreate a key

Playgroundsend a request

Tracesfind it again

Each stop is one page of the app, reached from the sidebar. You need one thing from outside Cloud Score: an API key from a model provider.

Any provider works for a first request. Cloud Score classifies requests with a small model and powers its semantic cache with embeddings from a connected OpenAI provider; without one, classification falls back to built-in rules and the semantic cache stays off.

Connect a provider

The Integrations page: one card per provider, connected ones with a status badge and health dots, unconnected bays with a Connect button. — Integrations: connected providers show their status and recent health checks; empty bays say what they need.

Click Providers under Gateway in the sidebar.
Every provider has a bay. An unconnected one is badged Not set up and says what it needs, for example needs API key.
On the OpenAI bay, click Connect.
The Connect OpenAI dialog opens.
Paste your key into API key *.
Leave Base URL at https://api.openai.com/v1 unless you route through a proxy.
Click Test.
Wait for Credentials look good. If it says Failed: …, fix the key and test again. Nothing is saved yet.
Click Save.
The card appears with the badge Disconnected: the credentials are stored (encrypted, shown as key •••• plus the last four characters) but no check has run yet.
Click Re-test on the card.
The badge flips to Connected and the first health dot appears. Within a minute Cloud Score adds a dot with a measured latency, and keeps checking from then on.

Disconnect has no confirmation.

Disconnect on a card deletes the provider, every model discovered from it, and its health history the moment you click.

Enable models

Routing only ever chooses a model that is enabled on this page. A fresh provider has none.

The Models page: a table of models with capabilities, context window, prices, quality and latency tiers, and an enable switch per row. — Models: the registry routing chooses from. Only enabled rows can be selected.

Click Models under Gateway in the sidebar.
Click Discover OpenAI models.
The button reads Discovering…, then a note such as Added 57 from LiteLLM (disabled — enable to route) · updated 57. Every discovered model arrives disabled, with its prices, context window, and capabilities filled in.
Set Provider to OpenAI.
The status line reads 0 of 57 enabled.
Tick one or two rows and click Enable selected.
The status line changes to 2 selected while ticking and the note reads Enabled 2 modelsafterwards. A row's own switch does the same for one model.

Azure AI Foundry has no discover button: its deployments are registered by your operator through the API.

Create an API key

Every request to the gateway, including from the Playground, carries a platform API key. The key decides the workspace, the security profile, and the budget a request runs under.

The Key created dialog on Admin › API keys showing a one-time dex_sk_ key with Copy and Done buttons. — Key created: the only time the full key is shown. Copy it before clicking Done.

Click API keys under Admin in the sidebar.
The key table opens. With no keys it reads No API keys yet — create one to call the gateway.
Click Create key.
Give it a Name, for example first-key.
Leave Store request/response bodies checked so Logs can show you the prompt and the answer. Leave Sandbox key — may hand-pick policies per request (playground) unchecked and Security profile on Observe (default).
Click Create.
The dialog becomes Key created and shows the full key, which starts with dex_sk_.
Click Copy, store the key somewhere safe, then click Done.
The full key is never shown again. The list shows only its prefix from now on.

Send a request

Click Playground in the sidebar.
Paste the key into Platform API key.
Switch on Save for session to keep it for this browser tab. It is cleared when you sign out or open the sign-in page.
Type a prompt in Message.
For example: Explain in two sentences what a routing gateway does.
Leave Routing on auto.
The caption under auto reads workspace rules decide: with no policies yet, Cloud Score uses its default balance of cost, latency, and quality. Stream the answer, under Advanced options, is on by default.
Click Optimize prompt, then choose a prompt.
With Prompt Optimization ON, Cloud Score prepares a bounded Balanced rewrite and records its API expense before generating your answer. Review both versions, the diff and estimated input savings. Choose Run original or Run optimized to start answer routing. Structural checks do not guarantee preserved meaning. With optimization OFF, Run original starts the request directly.

Read the result

The answer streams into the Result pane, alongside the Prompt Optimization receipt. Click Details: on Routing, stations appear from left to right as each stage finishes. Each station shows a one-line summary and how long the stage took.

Station	Its summary line tells you
`Guard in`	What AI Security did to the prompt: `off` when the key has no profile, otherwise the action and the categories it matched.
`Classify`	The task type and where the classification came from, for example `generation · heuristic` or `summarization · llm`.
`Match`	How many enabled models can do this task, for example `2 models`.
`Estimate`	The cheapest candidate's estimated cost, for example `from $0.000125`.
`Policy`	How many policies were evaluated, or `blocked`.
`Resolve`	The cache verdict. A first request reads `miss` or the reason it was not eligible.
`Select`	The winning model.
`Attempt` / `Execute`	The model called, then the measured cost, for example `$0.000118`. A failed attempt drops below the rail and the next `Select` is the failover.
`Guard out`	What AI Security did to the answer.
`Validate`	`ok`, or `retry` when the answer was empty or not the JSON you asked for.
`Record`	The final status: `completed`, `blocked`, or `failed`.

Under the stations, the decision summary names the model Cloud Score routed to with its score. Usage shows tokens, cost, and latency, and Decision holds the Optimization decision card, which lists every candidate with the cost difference each would have made. In Usage, click Copy ID to keep the request id, which starts with req_.

Same prompt, two strategies

What you do: Send the prompt with cost selected (caption cheapest within 8% of the best), then click intelligence (caption the best model, cheapest among ties) and Send again.
What Cloud Score does: Scores the same candidates twice with different cost, latency, and quality weights.
What you see: A different model in the result header, and in Details › Decision a Candidates table whose Δ vs selected column shows what each alternative would have cost.

Find it in Logs

Observability › Requests with the Request drawer open: the request id with Copy id, View full request and View in raw data, then Request details and the start of Metrics. — The Request drawer in Observability: the request's recorded details, and further down the drawer the same stations the Playground showed.

Click Traces in the sidebar.
Your request is the top row: time, the prompt preview, the model, and the status completed.
Click the row.
The Request detail drawer shows the same route board, the Optimization decision card, and the Cache, Classification, Request, and Response sections. Close it with Close details.
Look at Overview too.
Total requests and Request activity on the Default tab include your request in the selected period. If Overview was already open, use Refresh to update its snapshot. On Gateway insights, Evidence on Requests & spend opens the request evidence to inspect the recorded outcome.

What to do next

Point your application at Cloud Score from Admin › Developer, which generates a snippet for your language with the key and endpoint filled in. Then shape routing with Policies, cap spend under Budgets, and protect prompts and answers with Gateway › Guardrails.